Data analysis driven machining working hour intelligent estimation method and system

Abstract

This invention discloses a data analysis-driven intelligent prediction method and system for machining time, belonging to the field of industrial data processing technology. It includes: loading the 3D model data of the target part; performing multimodal feature recognition based on geometric topology information and surface visual information to obtain machining feature information; performing process decision-making and theoretical time calculation based on the machining feature information to obtain theoretical time and theoretical toolpath data; loading a digital twin of the target equipment and synchronizing equipment status data with the physical machining equipment in real time; inputting the theoretical toolpath data into the digital twin, performing virtual machining simulation to obtain the simulation time, and calculating the deviation between the theoretical time and the simulation time; and dynamically adjusting the time calibration coefficient through a reinforcement learning model based on the equipment status data, machining feature information, and deviation value to obtain the final predicted time. This invention effectively improves the accuracy and reliability of machining time prediction.

Description

Technical Field

[0001] This invention relates to the field of industrial data processing technology, specifically to a data analysis-driven intelligent prediction method and system for machining time. Background Technology

[0002] In the machining field, time estimation is the main basis for cost accounting and production scheduling. With the digital and intelligent upgrading of the manufacturing industry, as well as the expansion of enterprise business scale and the upgrading of order-taking models, the market has put forward higher requirements for the efficiency and accuracy of time estimation. Existing time estimation methods have gradually developed from manual experience estimation to automation.

[0003] However, existing methods mostly use static coefficients in the calculation process, failing to consider the real-time impact of dynamic equipment conditions such as machine tool aging, tool wear, and load fluctuations on processing time. The estimated results are prone to failure in actual production, resulting in insufficient accuracy and reliability of time estimation. In actual production, frequent manual intervention and adjustment are still required, making it difficult to give full play to the advantages of automated estimation. Summary of the Invention

[0004] This invention provides a data analysis-driven intelligent prediction method and system for machining time, aiming to solve the technical problem of insufficient accuracy and reliability of existing machining time prediction.

[0005] In view of the above problems, the present invention provides a data analysis-driven intelligent prediction method and system for machining time.

[0006] In a first aspect, the present invention provides a data analysis-driven intelligent prediction method for machining time, including:

[0007] Load the three-dimensional model data of the target part, wherein the three-dimensional model data includes geometric topology information and surface visual information;

[0008] Based on the geometric topology information and the surface visual information, multimodal feature recognition is performed to obtain processing feature information, wherein the processing feature information includes feature type, feature size and feature position;

[0009] Based on the machining feature information, process decision-making and theoretical time calculation are performed to obtain theoretical time and theoretical toolpath data;

[0010] A digital twin of the target device is loaded, wherein the digital twin includes a geometric model, a physical model and a behavioral model, and the device status data is synchronized with the physical processing device in real time, the device status data including spindle load data, tool wear data, vibration data and temperature data;

[0011] The theoretical toolpath data is input into the digital twin, virtual machining simulation is performed, simulation time is obtained, and the deviation between the theoretical time and the simulation time is calculated.

[0012] Based on the equipment status data, the processing feature information, and the deviation value, the time calibration coefficient is dynamically adjusted through a reinforcement learning model to obtain the final estimated working time.

[0013] Secondly, this invention provides a data analysis-driven intelligent prediction system for machining time, comprising:

[0014] A 3D model loading module is used to load the 3D model data of the target part, wherein the 3D model data includes geometric topology information and surface visual information;

[0015] A multimodal feature recognition module is used to perform multimodal feature recognition based on the geometric topology information and the surface visual information to obtain processing feature information, wherein the processing feature information includes feature type, feature size and feature position;

[0016] The process decision and time calculation module is used to perform process decision and theoretical time calculation based on the machining feature information to obtain theoretical time and theoretical toolpath data;

[0017] A digital twin loading module is used to load a digital twin of a target device. The digital twin includes a geometric model, a physical model, and a behavioral model, and synchronizes device status data with the physical processing device in real time. The device status data includes spindle load data, tool wear data, vibration data, and temperature data.

[0018] The virtual machining simulation module is used to input the theoretical toolpath data into the digital twin, perform virtual machining simulation, obtain the simulation time, and calculate the deviation between the theoretical time and the simulation time.

[0019] The intelligent time calibration module is used to dynamically adjust the time calibration coefficient based on the equipment status data, the processing feature information and the deviation value through a reinforcement learning model to obtain the final estimated working time.

[0020] One or more technical solutions provided in this invention have at least the following technical effects or advantages:

[0021] This invention provides a data analysis-driven intelligent prediction method and system for machining time: First, by loading a 3D model containing geometric topology and surface visual information and performing multimodal feature recognition, accurate extraction of part machining features is achieved, providing an accurate data foundation for subsequent calculations. Second, by combining process decisions and theoretical time calculations, theoretical toolpaths and machining times are obtained. Simultaneously, a digital twin synchronized with the physical equipment in real time is introduced. Through virtual machining simulation, real-time status data such as spindle load and tool wear are dynamically captured, and the deviation between theoretical and simulated machining times is calculated. Finally, based on equipment status, machining features, and deviation values, a reinforcement learning model is used to dynamically adjust the time calibration coefficient, achieving adaptive compensation for dynamic factors. This invention effectively solves the problem that traditional static coefficients cannot adapt to real-time changes such as machine tool aging and load fluctuations, improving the accuracy, intelligence, and automation level of time prediction, reducing manual intervention, and providing a reliable basis for production scheduling and cost accounting. Attached Figure Description

[0022] Figure 1 , Figure 2 A flowchart illustrating the data analysis-driven intelligent prediction method for machining time provided in this embodiment of the invention;

[0023] Figure 3 , Figure 4 This is a schematic diagram of the structure of a data analysis-driven intelligent prediction system for machining time provided in an embodiment of the present invention;

[0024] The components represented by each number in the attached diagram are explained below:

[0025] 11. 3D model loading module, 12. Multimodal feature recognition module, 13. Process decision and time calculation module, 14. Digital twin loading module, 15. Virtual machining simulation module, 16. Intelligent time calibration module. Detailed Implementation

[0026] This invention provides a data analysis-driven intelligent prediction method and system for machining time, which addresses the technical problem of insufficient accuracy and reliability in existing machining time prediction methods.

[0027] Example 1, as Figure 1 , Figure 2 As shown, this invention provides a data analysis-driven intelligent prediction method for machining time, the method comprising:

[0028] S100: Load the three-dimensional model data of the target part, wherein the three-dimensional model data includes geometric topology information and surface visual information.

[0029] In this embodiment of the invention, three-dimensional model data of the target part is loaded, wherein the three-dimensional model data includes geometric topology information and surface visual information. Machining feature recognition is a preliminary step in machining time estimation; the accuracy of machining feature recognition directly determines the accuracy of subsequent process decisions and time calculations. The foundation of machining feature recognition is comprehensive and accurate three-dimensional model data of the part. Geometric topology information reflects the structural and dimensional characteristics of the part and is key data for determining the type and specifications of machining features. Surface visual information supplements the spatial morphological characteristics of the part's surface, forming multimodal data complementarity with geometric topology information, improving the completeness and accuracy of subsequent feature recognition. However, in practical applications, the three-dimensional model files of parts may have geometric incompleteness, topological errors, etc., and direct parsing can easily lead to data extraction failure or distortion. Furthermore, the original model lacks a standardized surface visual data format, which cannot directly meet the needs of multimodal feature recognition. Therefore, a standardized process is needed to load the three-dimensional model data, complete model verification, geometric topology information extraction, and standardized construction of surface visual information, providing high-quality, multi-dimensional foundational data for subsequent multimodal feature recognition.

[0030] Step S100 in the method provided in this embodiment of the invention includes:

[0031] Obtain the solid 3D model file of the target part;

[0032] Perform geometric integrity verification on the entity 3D model file to obtain a verified entity 3D model.

[0033] The verified 3D model of the entity is parsed, and the geometric topology information is extracted, wherein the geometric topology information is represented in the form of an attribute adjacency graph;

[0034] The verified 3D entity model is rendered in 3D voxelization to generate voxel mesh data, which is then added to the surface visual information.

[0035] First, obtain the solid 3D model file of the target part. A solid 3D model file is a digital file constructed using CAD design software that fully expresses the 3D geometry and structure of the part; it serves as the carrier for storing the part's 3D model data. The solid 3D model file of the target part is received through a pre-defined model acquisition channel. This channel supports user-uploaded model files and can also retrieve the corresponding model file of the target part from a product design database. It natively adapts to mainstream solid 3D model formats such as STEP, STP, SLDPRT, PRT, CATPART, and X_T, and can directly interface with export files from various CAD design software.

[0036] For example, if the target part is a mold steel core of an automotive parts mold, the STEP format solid 3D model file of the core is exported from the UG design software, and the file is received through the preset model upload portal to obtain the solid 3D model file of the mold steel core.

[0037] Secondly, the geometric integrity of the entity 3D model file is verified to obtain a verified entity 3D model. Geometric integrity verification refers to inspecting the internal data of the entity 3D model file to verify whether the model has geometric incompleteness, topological errors, unmerged multiple bodies, dimensional overtravel, etc., to ensure that the model has a complete geometric structure that can be parsed and from which valid data can be extracted.

[0038] Specifically, the underlying data of the solid 3D model file undergoes a full-dimensional, element-by-element inspection, completing three checks in sequence: First, a single-body structure verification is performed to check for the existence of multiple independent solid units in the model data, ensuring that the model is a complete structure of a single machined part; second, geometric incompleteness detection is performed, traversing the underlying data of the model's faces, edges, and points to check for issues such as missing face data, missing endpoints in edge data, or completely overlapping coordinates of multiple point data; finally, topological error detection is performed, verifying the inherent connection relationships between faces and faces, and between edges and faces, to check for issues such as face boundary intersections and conflicting containment relationships between edges and faces. If any problem is found in any of the verification steps, the user is informed of the specific error type and prompted to re-upload a valid file; if all verification steps are successful, a geometrically complete solid 3D model is directly output.

[0039] For example, to perform geometric integrity verification on the STEP format file of a mold steel core, first, check the model data to confirm that the core is a single solid structure without independent redundant solid units; then, traverse the face, edge, and point data of the core's forming surface, mounting surface, and positioning holes to confirm that there are no missing face data, no overhanging edges in the positioning holes, and no overlapping points; finally, check the connection relationship between the forming surface and the mounting surface, and the inclusion relationship between the positioning hole edge and the hole surface to confirm that there are no topological errors such as face boundary intersections and relationship conflicts. After the verification is passed, the complete solid 3D model of the mold steel core is obtained.

[0040] Further, the verified 3D model of the entity is analyzed to extract the geometric topology information, which is represented in the form of an attribute adjacency graph. Geometric topology information refers to the underlying data in the 3D model that describes the geometric shape and structural relationships of the part. Geometric information refers to quantitative parameters such as the specific dimensions, positions, curvature, and normals of faces, edges, and points; topology information refers to the adjacency relationships between faces, the containment relationships between edges and faces, and the concavity and convexity of faces, etc. An attribute adjacency graph is a topology graph that structurally expresses the geometric topology information of the part, with faces, edges, and points as nodes, and the adjacency, containment, and connection relationships between nodes as edges, and each node is accompanied by corresponding geometric parameter attributes.

[0041] Specifically, the validated 3D solid model undergoes layered analysis. First, all faces, edges, and points of the model are traversed to extract the quantitative parameters corresponding to each structure, forming a geometric information dataset. Then, based on the overall structure of the model, the actual connection relationships between faces, edges, and points are analyzed, and features such as adjacency, inclusion, and concavity / convexity are identified, forming a topological information dataset. Finally, the geometric information dataset and the topological information dataset are merged. With faces, edges, and points as core nodes, each node is assigned corresponding geometric parameter attributes. Using the connection relationships between nodes as edges, an attribute adjacency graph is constructed to complete the extraction and structured storage of geometric and topological information.

[0042] For example, a layered analysis is performed on the verified 3D model of the mold steel core. First, the curvature of the core forming surface is R5mm, the size of the mounting surface is 100mm×80mm, the diameter of the positioning hole is 8mm and the center coordinates are (20,30,0), etc., to form a geometric information dataset. Then, the topological relationships such as the adjacent relationship between the forming surface and the mounting surface, the edge structure of the positioning hole being completely contained in the hole surface, and the outer contour surface of the core being a convex surface are sorted out to form a topological information dataset. Finally, the forming surface, mounting surface, positioning hole, etc. are used as nodes of the attribute adjacency graph, and the above geometric parameters are attached to each node. The nodes are connected by adjacency and containment relationships as edges to generate the attribute adjacency graph of the core, thus completing the extraction and storage of geometric topological information.

[0043] Next, the verified 3D solid model is subjected to 3D voxel rendering to generate voxel mesh data, which is then added to the surface visual information. 3D voxel rendering refers to discretizing a continuous 3D solid model into regularly arranged cubic voxel units, and digitally reconstructing the 3D surface morphology of the part by assigning feature values ​​to each voxel unit. Voxel mesh data refers to the dataset obtained after 3D voxel rendering, containing surface visual feature parameters such as the spatial coordinates, grayscale value, and voxel density of each voxel unit. Surface visual information refers to the dataset describing the spatial morphology and visual features of the part's surface, and is one of the important data dimensions for multimodal feature recognition.

[0044] Specifically, based on the machining accuracy requirements of the parts, a matching voxel resolution is set. The higher the machining accuracy requirement, the smaller the voxel resolution. Based on the set voxel resolution, a 3D voxelization rendering operation is performed on the verified solid 3D model. The 3D surface of the model is discretized into regular cubic voxel units. All voxel units are traversed, and feature parameters such as spatial coordinates and grayscale values ​​of each voxel unit are extracted to form standardized voxel mesh data. This voxel mesh data is then incorporated into the surface visual information dataset of the target part to complete the construction of surface visual information.

[0045] For example, for a solid 3D model of a mold steel core, the required machining accuracy is 0.1mm, and a voxel resolution of 0.1mm is set accordingly. Based on this resolution, 3D voxel rendering is performed on the core, and all surface structures of the core, such as the forming surface, the inner wall of the positioning hole, and the outer contour, are discretized into cubic voxel units of 0.1mm×0.1mm×0.1mm. All voxel units are traversed, and the spatial coordinates and corresponding gray values ​​of each voxel unit are extracted to generate voxel mesh data containing the complete surface morphology of the core. This data is added to the surface visual information of the mold steel core to complete the construction of the surface visual information.

[0046] In this embodiment of the invention, a multi-format compatible model acquisition method is adopted to adapt to the model output formats of various CAD design software in practical applications, thereby improving the compatibility and practical applicability of the method. Geometric integrity verification effectively eliminates invalid models with geometric incompleteness or topological errors, ensuring the effectiveness of subsequent data extraction and feature recognition from the data source and avoiding errors in subsequent calculations due to problems with the model itself. Structured parsing extracts geometric topological information carried by attribute adjacency graphs, achieving standardized and structured storage of part structure and dimensional data, facilitating rapid retrieval and analysis in subsequent feature recognition stages. Three-dimensional voxel rendering generates voxel mesh data and supplements it to surface visual information, realizing the digital reconstruction of the part's surface morphology and providing complete visual data support for multimodal feature recognition.

[0047] S200: Based on the geometric topology information and the surface visual information, perform multimodal feature recognition to obtain processing feature information, wherein the processing feature information includes feature type, feature size, and feature position. Processing feature information is the direct basis for subsequent process decisions and theoretical time calculations; its precision and accuracy directly determine the final effect of time estimation. Fine-grained processing features are a refined breakdown of coarse-grained features, directly corresponding to specific processing techniques. Combining multimodal recognition with geometric topology information and surface visual information, multidimensional and accurate feature determination can be achieved. Therefore, it is necessary to complete targeted training of the fine-grained feature recognition model using historical processed part data, construct a model set classified by part type, and then call the matching model for the target part to complete multimodal feature recognition, outputting fine-grained processing feature information including type, size, and position, providing an accurate basis for subsequent process and time calculations.

[0048] Step S200 in the method provided in this embodiment of the invention includes:

[0049] Collect several historically processed parts of a preset part model. For each historically processed part, extract its geometric topology record data and surface visual record data. At the same time, label the first fine-grained feature identifier up to the Pth fine-grained feature identifier contained in the historically processed part, where P is the total number of preset fine-grained processing feature categories. The fine-grained processing features include through hole features, blind hole features, countersunk hole features, U-groove features, T-groove features, dovetail groove features, rectangular cavity features, circular cavity features, and chamfer features.

[0050] Using the first fine-grained feature identifier as supervision and the geometric topology recording data and the surface visual recording data as input, a first feature recognizer capable of recognizing the first fine-grained feature is trained.

[0051] The process continues until the Pth fine-grained feature identifier is used as supervision, and the geometric topology recording data and the surface visual recording data are used as inputs to train a Pth feature recognizer capable of recognizing the Pth fine-grained features.

[0052] Merge the output layers of the first feature recognizer up to the Pth feature recognizer to construct a multi-label classification model, obtain a multi-modal feature recognition model, store it in association with the preset part model, and add it to the multi-modal feature recognition model set;

[0053] Based on the target part model, the multimodal feature recognition model set is invoked, and the geometric topology information and surface visual information of the current part are input into the multimodal feature recognition model. Through forward propagation, the probability values ​​of the current part belonging to the first processing feature up to the Pth processing feature are obtained simultaneously. The processing features with a probability value greater than a preset probability threshold are taken as the recognition result, and the processing feature information is output.

[0054] First, collect several historically processed parts of a preset part model. For each historically processed part, extract its geometric topology record data and surface visual record data, and simultaneously label the first fine-grained feature identifier up to the Pth fine-grained feature identifier contained in the historically processed part, where P is the total number of preset fine-grained processing feature categories. The fine-grained processing features include through hole features, blind hole features, countersunk hole features, U-groove features, T-groove features, dovetail groove features, rectangular cavity features, circular cavity features, and chamfer features.

[0055] Among them, the preset part model refers to the part category divided according to the part's structural type and processing technology attributes. Parts of the same model have similar processing features and process requirements. Geometric topology record data refers to the geometric topology information extracted from historically processed parts before processing, which is a record of the underlying data such as the structure and dimensions of historical parts. Surface visual record data refers to the surface visual information extracted from historically processed parts before processing, which is a digital record of the surface morphology of historical parts. Fine-grained feature identifier refers to the exclusive identifier set for each type of fine-grained processing feature, used to mark the feature type of historical parts. Fine-grained processing features are specific feature categories with clear processing technology meanings obtained by refining traditional coarse-grained features, such as subdividing holes into through holes, blind holes, countersunk holes, etc. P is the total number of preset fine-grained processing feature categories. In this embodiment, P=9, corresponding to 9 types of fine-grained processing features: through holes, blind holes, countersunk holes, U-grooves, T-grooves, dovetail grooves, rectangular cavities, circular cavities, and chamfers.

[0056] Specifically, first, determine the preset part model to which the part to be identified belongs, and collect several historically processed parts of the same model; for each historically processed part, extract its geometric topology record data and surface visual record data before processing to form a one-to-one corresponding two-dimensional data set; then, compare it with 9 types of fine-grained processing features, perform feature verification on each historically processed part, and label each type of fine-grained processing feature it actually contains with the corresponding fine-grained feature identifier, thus completing the labeling process of historical data.

[0057] For example, the target part is determined to be a mold steel core, and its corresponding preset part model is a mold core type part; 2000 historically processed mold steel core parts are collected according to this model. For each core, its geometric topology record data and surface visual record data are extracted to form 2000 sets of two-dimensional data; feature verification is performed on one of the cores, and it is found that it contains three types of fine-grained features: blind holes, chamfers, and rectangular cavities. The core is labeled with blind hole feature identifiers, chamfer feature identifiers, and rectangular cavity feature identifiers. The feature labeling of 2000 historical cores is completed in this way.

[0058] Secondly, using the first fine-grained feature identifier as supervision and the geometric topology recording data and the surface visual recording data as input, a first feature recognizer capable of recognizing the first fine-grained feature is trained.

[0059] Again, the P-th feature recognizer is trained using the P-th fine-grained feature identifier as supervision and the geometric topology recording data and the surface visual recording data as input, until the P-th fine-grained feature is recognized.

[0060] The feature recognizer is a recognition carrier used to determine whether a part contains a certain type of fine-grained processing feature. It takes geometric topology recording data and surface visual recording data as input, and is trained with the corresponding fine-grained feature identifier as supervision. After training, it can output the probability value of the fine-grained feature.

[0061] Specifically, the feature recognizer is trained sequentially according to the order of the first to ninth fine-grained feature labels. For each type of fine-grained feature, a binary classification recognition model with a uniform structure is built as the dedicated recognizer for that type of feature. The recognition model has three hidden layers: the first hidden layer has 64 neurons, the second hidden layer has 32 neurons, and the third hidden layer has 16 neurons. The hidden layers all use the ReLU activation function to achieve non-linear feature mapping. The output layer has one neuron and uses the Sigmoid activation function to output the probability value.

[0062] During training, the geometric topology records of historically processed parts are first converted into 1024-dimensional attribute adjacency graph feature vectors, and the surface visual records are reduced to 1024-dimensional voxel grid feature vectors. The two types of vectors are concatenated into a 2048-dimensional fused feature vector as the input to the recognition model. The fine-grained feature labels corresponding to this type of feature serve as the supervision basis—if the historical part contains this type of feature, it is labeled as 1, and if it does not, it is labeled as 0. This binary labeling result serves as the output label of the recognition model. The training adopts the mini-batch gradient descent method, with a batch size of 32 and an initial learning rate of 0.001. The learning rate is decayed using a cosine annealing strategy. The AdamW optimizer is selected, and the binary cross-entropy loss function is used to calculate the error between the model's predicted value and the true label value.

[0063] Secondly, during the training process, 2000 historical data points were divided into a training set and a validation set in a 7:3 ratio. After each round of training, the model performance was verified on the validation set. The convergence condition was set as follows: the number of training rounds did not exceed 100 rounds, the validation set loss did not decrease for 10 consecutive rounds, and the validation set recognition accuracy was ≥98%. Training of this type of feature recognizer was stopped when any convergence condition was met.

[0064] For example, for blind hole features, a binary classification recognition model with the aforementioned three hidden layers is constructed as a blind hole feature recognizer. The geometric topological record data and surface visual record data of 2000 historical mold steel core parts are converted into 1024-dimensional feature vectors and concatenated into a 2048-dimensional fused feature vector, which serves as the input data for the blind hole feature recognizer. Simultaneously, the 2000 historical core parts are binary-labeled for blind hole features: cores containing blind hole features are labeled as 1, and those not containing them are labeled as 0, serving as the output labels. Training is conducted using an AdamW optimizer with a batch size of 32 and an initial learning rate of 0.001, along with a cosine annealing learning rate decay strategy, with the binary cross-entropy loss function as the optimization objective. The training and validation sets were divided in a 7:3 ratio. Training was stopped when the validation set loss showed no decrease for 10 consecutive rounds and the validation set recognition accuracy reached 98.5%, thus completing the training of the blind hole feature recognizer. Using the same model structure, input / output format, training parameters, loss function, and convergence condition, fine-grained feature identifiers for chamfers, rectangular cavities, through holes, countersunk holes, U-grooves, T-grooves, dovetail grooves, and circular cavities were used as supervision to conduct targeted training on two-dimensional data of 2000 historical cores. Each feature recognizer was trained until it met the convergence condition, resulting in 9 feature recognizers that could independently identify the corresponding fine-grained features. The validation set recognition accuracy of each recognizer was ≥98%.

[0065] Furthermore, the output layers of the first feature recognizer up to the Pth feature recognizer are merged to construct a multi-label classification model, obtaining a multimodal feature recognition model. This model is then associated with and stored for the preset part model and added to the multimodal feature recognition model set. The multi-label classification model is a recognition carrier that can simultaneously determine whether a part contains multiple fine-grained processing features, unlike a single-label recognition carrier which can only determine one feature. It is suitable for scenarios where multiple features coexist in actual part processing. The multimodal feature recognition model is a model that combines geometric topology and surface visual dual-dimensional data to achieve fine-grained multi-feature recognition. It is derived from the multi-label classification model and is a dedicated recognition model adapted to a specific preset part model. The multimodal feature recognition model set consists of multimodal feature recognition models corresponding to different preset part models, stored according to part model, and can be retrieved as needed based on the target part model.

[0066] Specifically, the output layers of the nine trained feature recognizers are merged and combined to construct a multi-label classification model that can simultaneously output nine types of fine-grained feature recognition results. This model is a multi-modal feature recognition model adapted to mold core parts. A unique association is established between this multi-modal feature recognition model and the part model, and then it is stored in the multi-modal feature recognition model set and classified and stored according to the part model.

[0067] For example, the output layers of nine mold steel core feature recognizers, such as blind holes, chamfers, rectangular cavities, and through holes, are merged and combined to construct a multi-label classification model. This model can simultaneously identify whether a part contains the above nine types of fine-grained features, thus obtaining a multi-modal feature recognition model for mold core parts. This model is then bound to the model number of the mold core parts and stored in the multi-modal feature recognition model set.

[0068] Finally, based on the target part model, the multimodal feature recognition model set is invoked, and the geometric topology information and surface visual information of the current part are input into the multimodal feature recognition model. Through forward propagation, the probability values ​​of the current part belonging to the first processing feature up to the Pth processing feature are obtained simultaneously. The processing features with a probability value greater than a preset probability threshold are taken as the recognition result, and the processing feature information is output.

[0069] The target part model refers to the preset part model to which the current part belongs, used to match the corresponding multimodal feature recognition model. The geometric topology information and surface visual information of the current part refer to the two-dimensional basic data of the part to be processed, loaded and extracted from S100, serving as input data for feature recognition. Forward propagation refers to the process of sequentially calculating the input geometric topology information and surface visual information through the layers of the multimodal feature recognition model, ultimately outputting the recognition probability values ​​of each fine-grained feature. The preset probability threshold is a critical value set for the credibility of the feature recognition result; if the probability value is higher than this threshold, it is determined that the part actually contains the corresponding fine-grained feature. In this embodiment, the preset probability threshold is set to 90%. The processing feature information includes complete information such as the type of fine-grained processing feature actually possessed by the current part, the specific dimensions of each feature, and the spatial position of each feature.

[0070] Specifically, first, the target part model corresponding to the current part to be processed is determined. Based on this model, a multimodal feature recognition model with a unique association is retrieved from the multimodal feature recognition model set. Then, the geometric topology information and surface visual information of the current part extracted from S100 are synchronously input into the retrieved multimodal feature recognition model. The multimodal feature recognition model performs forward propagation calculation to obtain the probability values ​​of the current part belonging to 9 types of fine-grained processing features. The probability value of each type of feature is compared with a preset probability threshold, and fine-grained features with probability values ​​greater than the threshold are selected as the final feature recognition result of the current part. The specific size and spatial position of each identified feature are extracted from the geometric topology information, integrated with the feature type into complete processing feature information, and output.

[0071] For example, the target part model corresponding to the mold steel core to be processed is determined to be a mold core type part. Based on this model, the corresponding mold core type multimodal feature recognition model is retrieved from the multimodal feature recognition model set. The geometric topology information and surface visual information of the mold steel core extracted in S100 are simultaneously input into the multimodal feature recognition model. Through forward propagation calculation, the probability values ​​of the nine fine-grained features corresponding to the core are obtained, including 98% for blind hole features, 95% for chamfer features, and 95% for rectangular cavity features. 96% of the features were identified as blind holes, chamfers, and rectangular cavities, while the probabilities of the other six features were all below 90%. Based on a preset probability threshold of 90%, blind holes, chamfers, and rectangular cavities were selected as the identification results for the core. From the geometric topology information, the following were extracted: blind hole diameter 8mm, spatial center coordinates (20,30,0), chamfer angle 45°, side length 5mm, located at the edge of the core mounting surface, and rectangular cavity size 30mm×20mm×10mm, located at the center of the core forming surface. The feature type, size, and position were integrated to output the processing feature information of the mold steel core.

[0072] In this embodiment of the invention, by training a feature recognizer separately for each type of fine-grained feature, interference from multi-feature mixed training is avoided, thus improving the accuracy of single feature recognition. A multi-label classification model is constructed by merging the output layers, enabling simultaneous recognition of multiple fine-grained features and adapting to real-world scenarios where multiple features coexist in machined parts. Simultaneously, a multi-modal recognition method combining geometric topology information and surface visual information compensates for the information loss problem in single-dimensional feature recognition. By filtering recognition results through preset probability thresholds, the reliability and accuracy of feature recognition are further improved. The final output, containing fine-grained machining feature information including feature type, feature size, and feature location, directly corresponds to specific machining processes, providing accurate and specific basis for subsequent process decisions and theoretical time calculations, reducing errors in subsequent time calculations.

[0073] S300: Based on the machining feature information, perform process decision-making and theoretical time calculation to obtain theoretical time and theoretical toolpath data.

[0074] In this embodiment of the invention, based on the machining feature information, process decision-making and theoretical time calculation are performed to obtain theoretical time and theoretical toolpath data. After the machining feature information is determined, a reasonable machining process, matching appropriate tools and cutting parameters, must be planned through scientific process decision-making to generate theoretical toolpath data that conforms to the actual machining logic, thereby completing accurate theoretical time calculation. The process knowledge graph, as a structured carrier of process knowledge, can achieve intelligent matching of machining features with historical process schemes. The preset cutting parameter library provides standardized and authoritative cutting parameters. Combined with the accurate overlay of auxiliary time, it can effectively solve the subjectivity problem of traditional process decision-making and improve the rationality of process planning and the accuracy of time calculation. Therefore, process decision-making and theoretical time calculation need to be completed through processes such as loading the process knowledge graph, matching historical process schemes, generating theoretical toolpaths, and overlaying auxiliary time, providing reliable process data support for subsequent digital twin simulation.

[0075] Step S300 in the method provided in this embodiment of the invention includes:

[0076] Load a process knowledge graph, wherein the process knowledge graph constructs a semantic network using part feature nodes, machining process nodes, tool nodes, and material nodes;

[0077] The processing feature information is input into the process knowledge graph, and similar historical parts are matched to obtain the initial process chain.

[0078] Based on the initial process chain and the preset cutting parameter library, theoretical toolpath data is generated, wherein the theoretical toolpath data includes the toolpath length and feed rate of each process;

[0079] Based on the theoretical toolpath data, the theoretical machining time is calculated and a preset auxiliary time is added to obtain the theoretical machining time.

[0080] First, a process knowledge graph is loaded. This graph constructs a semantic network using part feature nodes, machining process nodes, tool nodes, and material nodes. The process knowledge graph is a knowledge base that stores machining process knowledge in a structured semantic network format. It includes part feature nodes, machining process nodes, tool nodes, and material nodes, all connected by semantic relationships. It comprehensively records the associated process information, such as standard machining processes, suitable tool models, recommended cutting parameters, and material properties, corresponding to different part features. A semantic network is a knowledge representation form centered on nodes and linked by relationships. It clearly presents the correspondence between part features and machining processes, tools, and materials, enabling structured association and rapid retrieval of process knowledge.

[0081] Specifically, a process knowledge graph is pre-constructed and stored. During construction, mainstream fine-grained machining features in the machining field, such as through holes, blind holes, and chamfers, are first identified as part feature nodes. For each feature node, standardized machining operations, such as drilling, milling, and chamfering, are matched as machining operation nodes. Suitable tool types, such as twist drills, end mills, and chamfering cutters, are configured for each operation node as tool nodes, and core tool parameters, such as tool diameter, cutting edge length, and material, are associated. Simultaneously, the corresponding material types, such as mold steel, aluminum alloy, and cast iron, are recorded as material nodes, and parameters such as material hardness and machining characteristics are associated. Finally, semantic relationships are constructed: part feature node - containing feature - machining operation node - suitable tool - material node, forming a complete process knowledge graph.

[0082] For example, a process knowledge graph corresponding to the mold steel core is constructed. The blind hole feature is used as the part feature node. The drilling process is associated through the inclusion feature relationship, the φ8mm carbide twist drill is associated through the matching tool relationship, and the mold steel material is associated through the corresponding material relationship. At the same time, the chamfer feature node is associated with the chamfering process, the φ5mm 45° chamfering tool and the mold steel material. In this way, the process knowledge graph corresponding to the nine types of fine-grained features is constructed to form a semantic network.

[0083] Secondly, the machining feature information is input into the process knowledge graph to match similar historical parts and obtain an initial process chain. Machining feature information refers to the complete feature data output in S200, including fine-grained feature types, feature dimensions, and feature positions. Similar historical parts refer to historically machined parts that are consistent with the target part in terms of feature type, feature size, and material type; their corresponding process schemes have high reference value. The initial process chain refers to a standardized sequence of machining operations arranged in machining order based on the matching of similar historical parts. It includes the feature type, tool type, and basic cutting parameters of each operation and forms the basic framework for generating theoretical toolpaths.

[0084] Specifically, the target part processing feature information obtained in S200 is semantically matched with the part feature nodes in the process knowledge graph to filter out similar historical parts with the same feature type, size, and material; the complete processing sequence corresponding to the similar historical parts is retrieved from the process knowledge graph and organized into an initial process chain according to the processing sequence to clarify the processing object, suitable tool and basic process attributes of each process.

[0085] For example, the processing feature information of the mold steel core is used as input: blind hole feature φ8mm, chamfer feature 45° φ5mm, rectangular cavity feature 30mm×20mm×10mm, and the material is mold steel; similar historical parts with the same features are matched in the process knowledge graph, and the processing sequence of the historical part is retrieved: first, the drilling process is performed to process the blind hole, then the chamfering process is performed to process the edge of the mounting surface, and finally the milling process is performed to process the rectangular cavity. The initial process chain is organized according to the processing sequence, and the corresponding features, suitable tools and basic process attributes of each process are clarified.

[0086] Next, based on the initial process chain and the preset cutting parameter library, theoretical toolpath data is generated. This theoretical toolpath data includes the toolpath length and feed rate for each process step. The preset cutting parameter library refers to a pre-stored set of standardized machining parameters, containing recommended cutting parameters for different materials, tools, and characteristics, such as spindle speed, feed rate, and depth of cut. The data is derived from industry cutting manuals, tool manufacturer recommendations, and the client's historical machining experience, providing authoritative basis for process decisions. The theoretical toolpath data refers to standardized data describing the tool machining path generated based on the initial process chain and cutting parameters. It includes the toolpath length and feed rate for each process step and serves as input data for virtual machining simulation.

[0087] Specifically, the process iterates through each machining operation in the initial process chain. Based on the feature type, suitable tool model, and target part material of the operation, it retrieves the matching recommended cutting parameters from the preset cutting parameter library: spindle speed n, feed rate f, and depth of cut ap. It then calculates the toolpath length for each operation in conjunction with the feature dimensions. For example, the toolpath length for drilling is the blind hole depth, and the toolpath length for milling is the cavity contour perimeter. Finally, it associates the retrieved feed rate with the calculated toolpath length to form theoretical toolpath data that includes the toolpath length and feed rate of each operation.

[0088] For example, the initial process chain of the mold steel core is traversed. For the drilling process, recommended parameters are retrieved from the preset cutting parameter library: spindle speed n=1000r / min, feed rate f=60mm / min. Combined with the blind hole depth of 10mm, the toolpath length is calculated to be 10mm, and the toolpath length and feed rate of 60mm / min for this process are recorded. For the chamfering process, the recommended feed rate f=200mm / min is retrieved. Combined with the total chamfer edge length of 20mm, the toolpath length is calculated to be 20mm, and the corresponding data is recorded. For the milling of the rectangular cavity process, the recommended feed rate f=180mm / min is retrieved. Combined with the cavity contour perimeter of 100mm, the toolpath length is calculated to be 100mm. Finally, the toolpath lengths and feed rates of the three processes are integrated to generate theoretical toolpath data.

[0089] Furthermore, based on the theoretical toolpath data, the theoretical machining time is calculated and superimposed with a preset auxiliary time to obtain the theoretical machining time. The theoretical machining time refers only to the time the tool actually participates in cutting, obtained by summing the toolpath lengths of each process divided by the corresponding feed rates; it represents the time consumed by the cutting process itself. The preset auxiliary time refers to the time consumed by auxiliary operations other than cutting, including tool change time, idle travel time, workpiece clamping and unclamping time, tool rapid approach and retraction time, spindle acceleration and deceleration time, and waiting time between processes; these are standardized preset values. The theoretical machining time is the sum of the theoretical machining time and the preset auxiliary time, fully covering the entire machining process.

[0090] Specifically, first, calculate the theoretical processing time for all processes according to the formula: Theoretical processing time = Σ (toolpath length of each process / corresponding feed rate); second, sum up the components of the preset auxiliary time to obtain the total auxiliary time; finally, add them together to obtain the final theoretical working time: Theoretical working time = Theoretical processing time + Total auxiliary time.

[0091] For example, the theoretical machining time for the mold steel core is calculated as follows: Drilling time = 10 / 60 ≈ 0.167 min, chamfering time = 20 / 200 = 0.1 min, milling cavity time = 100 / 180 ≈ 0.556 min, total theoretical machining time = 0.167 + 0.1 + 0.556 = 0.823 min; summarizing the preset auxiliary times: tool change time 0.5 min, idle stroke 0.3 min, clamping and disassembly 0.8 min, rapid approach and retraction 0.2 min, spindle acceleration and deceleration 0.1 min, process waiting 0.1 min, total auxiliary time = 0.5 + 0.3 + 0.8 + 0.2 + 0.1 + 0.1 = 2 min; theoretical machining time = 0.823 + 2 = 2.823 min.

[0092] In this embodiment of the invention, by loading a structured process knowledge graph, intelligent and precise matching of machining features and historical process schemes is achieved, avoiding the subjectivity and arbitrariness of manual process decision-making and ensuring that process planning conforms to actual machining logic. Combined with a preset cutting parameter library, standardized and authoritative cutting parameters are matched to each process, ensuring the rationality and practicality of theoretical toolpath data and accurately reproducing the actual machining path and movement speed of the tool. Simultaneously, by splitting theoretical machining time and superimposing preset auxiliary time, the entire machining process time is fully covered, solving the core defect of traditional time calculation that only focuses on cutting time and ignores auxiliary operation time, thus achieving accurate calculation of theoretical time. The final output theoretical toolpath data provides accurate process input for subsequent digital twin virtual machining simulation, improving the scientificity and accuracy of the entire time estimation process.

[0093] S400: Load a digital twin of the target device, wherein the digital twin includes a geometric model, a physical model and a behavioral model, and synchronizes device status data with the physical processing device in real time, the device status data including spindle load data, tool wear data, vibration data and temperature data.

[0094] In this embodiment of the invention, a digital twin of the target device is loaded. The digital twin includes a geometric model, a physical model, and a behavioral model, and its device status data is synchronized with the physical machining equipment in real time. This device status data includes spindle load data, tool wear data, vibration data, and temperature data. The fidelity and state synchronization of the digital twin with the physical machining equipment directly determine the realism of the virtual machining simulation results, thus affecting the accuracy of the deviation calculation between theoretical and simulated machining times. Therefore, it is necessary to construct a three-in-one digital twin comprising a geometric model, a physical model, and a behavioral model to achieve a comprehensive and accurate mapping of the structure, characteristics, and motion behavior of the physical machining equipment. Furthermore, real-time synchronization of device status data is achieved through industrial communication protocols, ensuring that the operating state of the digital twin remains consistent with the physical equipment, thus providing an accurate and real-time virtual equipment carrier for subsequent virtual machining simulations.

[0095] Step S400 in the method provided in this embodiment of the invention includes:

[0096] A geometric model is established for the physical processing equipment, wherein the geometric model includes the three-dimensional structure of the machine tool bed, spindle, tool magazine and worktable;

[0097] A physical model is established for the physical processing equipment, wherein the physical model includes dynamic characteristic parameters and tool wear evolution law;

[0098] A behavioral model is established for the physical processing equipment, wherein the behavioral model simulates the interpolation logic and acceleration / deceleration control of the CNC controller;

[0099] The device status data is collected in real time using industrial communication protocols, and the status variables of the digital twin are updated synchronously.

[0100] First, a geometric model is established for the physical machining equipment. This geometric model includes the three-dimensional structure of the machine tool bed, spindle, tool magazine, and worktable. The geometric model is a 1:1 three-dimensional digital model constructed according to the actual physical parameters of the physical machining equipment. It serves as the basic carrier of the digital twin, used to reproduce the physical structure of the equipment, the relative positions of its components, and their kinematic relationships. The machine tool bed is the fundamental supporting component of the physical machining equipment, providing an installation reference for other components. Its geometric structure determines the overall layout of the equipment. The spindle is the moving component that drives the tool to rotate and achieve cutting. Its three-dimensional structure includes key parts such as the spindle head and tool clamping mechanism. The tool magazine is the component used to store various types of tools required for machining. Its geometric structure includes tool positions and tool changing mechanisms. The worktable is the component used to clamp the workpiece. Its three-dimensional structure includes the clamping surface and the motion slide, and it is crucial for realizing the spatial movement of the workpiece.

[0101] Specifically, the process begins by collecting a complete set of design drawings and actual measured dimensions of the physical processing equipment to obtain precise external dimensions, assembly dimensions, and kinematic pair parameters of the machine tool bed, spindle, tool magazine, and worktable. Then, using 3D modeling, the 3D solid structure of each component is recreated at a 1:1 scale, strictly adhering to the actual assembly relationships of the physical equipment, to build a spatial position model of each component. Finally, matching kinematic pair attributes are configured for each moving component to recreate the rotational motion of the spindle, the linear movement of the worktable, and the tool change linkage between the tool magazine and the spindle, forming a complete geometric model of the equipment.

[0102] For example, for a 3-axis CNC milling machine that processes mold steel cores, the design drawings and actual measured dimensions of the machine are collected to obtain the precise dimensions of its cast iron bed, electric spindle, 24-position disc-type tool magazine, and cross slide table. A 3D model of each component is constructed at a 1:1 scale. The electric spindle is assembled on the bed column, the worktable is assembled on the slide rail of the bed base, and the tool magazine is assembled in a preset position on the side of the bed. Rotary kinematic pairs are configured for the electric spindle, XYZ three-axis linear kinematic pairs are configured for the worktable, and rotary and traverse linkage kinematic pairs are configured for the tool magazine tool changer arm. The actual kinematic relationship of each component is restored to form the geometric model of the 3-axis CNC milling machine.

[0103] Secondly, a physical model is established for the physical machining equipment. This physical model includes dynamic characteristic parameters and tool wear evolution laws. The physical model is a digital model used to reproduce the dynamic characteristics of the physical machining equipment and the tool wear laws, reflecting the equipment's response under cutting forces and the influence of tool condition on machining. The dynamic characteristic parameters include the stiffness coefficients, natural frequencies, damping ratios, and spindle power-torque characteristic curves of each axis of the machine tool, used to simulate the machine tool's response and deformation under cutting forces. The tool wear evolution law refers to the quantitative law of tool wear variation with cutting time, cutting parameters, and machining materials, used to evaluate the impact of the actual tool condition on cutting efficiency and machining path.

[0104] Specifically, firstly, by using the technical manuals, modal tests, and performance tests provided by the equipment manufacturer, the stiffness coefficients, natural frequencies, damping ratios of each axis of the physical machining equipment (XYZ), as well as the power-torque characteristic curves of the spindle at different speeds, are obtained. These parameters are then integrated into a digital model to construct a machine tool dynamics model. Next, through extensive cutting tests and statistical analysis of historical machining data, a quantitative curve showing the change of tool wear over cutting time is fitted for different cutting tools, different machining materials, and different cutting parameters, forming a tool wear evolution law. Finally, the machine tool dynamics model and the tool wear evolution law are integrated to obtain a complete physical model of the equipment.

[0105] For example, for the aforementioned 3-axis CNC milling machine, its X-axis stiffness coefficient (80 N / μm), Y-axis stiffness coefficient (75 N / μm), and Z-axis stiffness coefficient (90 N / μm) were obtained through the manufacturer's technical manual and modal tests. The natural frequency of each axis was 50 Hz, the damping ratio was 0.05, and the power-torque characteristic curves at a spindle speed of 3000 r / min (5.5 kW power, 17.5 Nm torque) were used to construct a machine tool dynamics model. Then, through cutting tests of mold steel using carbide tools, the wear evolution law of machining mold steel with a φ8mm carbide twist drill was fitted: VB = 0.01t + 0.01, where VB is the width of the wear band on the tool's flank face (mm); t is the actual cutting time (min). This law, along with the wear laws of other machining tools, was incorporated into the model to form the physical model of the 3-axis CNC milling machine.

[0106] Next, a behavior model is established for the physical processing equipment, where the behavior model simulates the interpolation logic and acceleration / deceleration control of the numerical control (NC) controller. The behavior model refers to a digital model used to replicate the motion control logic of the NC controller of the physical processing equipment, which is the key for the digital twin to achieve consistent motion timing and trajectory with the physical equipment, and can convert the theoretical tool path into actual equipment motion commands that conform to the reality. The interpolation logic of the NC controller refers to the logic by which the numerical control system generates continuous motion commands for each axis based on the theoretical tool path, including trajectory generation methods such as linear interpolation, circular interpolation, and helical interpolation. The acceleration / deceleration control refers to a motion speed control strategy based on acceleration and jerk limits, which is used to control the start, acceleration, constant speed, deceleration, and stop processes of each axis of the equipment, avoid motion impacts, and ensure trajectory accuracy.

[0107] Specifically, first replicate the native interpolation logic of the NC controller supporting the physical processing equipment, and build an interpolation model that includes mainstream trajectory generation methods such as linear interpolation and circular interpolation, ensuring that the model can discretely generate segmented motion commands for each axis of the machine tool according to the continuous theoretical tool path; then, according to the motion performance parameters of the physical equipment, set the matching acceleration and jerk limit thresholds, and build an acceleration / deceleration control model to control the speed change process of each axis of motion; finally, integrate the interpolation model with the acceleration / deceleration control model, enabling the model to discretize the theoretical tool path into actual motion commands for each axis of the machine tool and strictly match the motion timing of the physical equipment, forming a complete equipment behavior model.

[0108] Exemplarily, for the above-mentioned 3-axis NC milling equipment with a supporting FANUC 0i-MF NC controller, replicate the linear interpolation and circular interpolation logics of the controller, build an interpolation model, which can discretely generate segmented motion commands for the X, Y, and Z axes according to the theoretical tool paths of milling blind holes and rectangular cavities; according to the motion performance of the equipment, set the jerk limit to 0.5g and the acceleration limit to 1g, where g is the acceleration due to gravity, build an acceleration / deceleration control model to control the start-up acceleration, cutting constant speed, and in-position deceleration processes of the worktable and the spindle; integrate the interpolation model with the acceleration / deceleration control model to form the behavior model of the equipment, which can convert the theoretical tool path into time-sequenced commands that conform to the actual motion of the equipment, ensuring that the time axis of virtual machining is consistent with the execution process of the physical equipment.

[0109] Finally, the equipment status data is collected in real time using an industrial communication protocol, and the status variables of the digital twin are updated synchronously. An industrial communication protocol is a standardized communication specification used for data transmission between industrial devices and between industrial devices and digital models; it is the foundation for real-time data acquisition and transmission. Equipment status data refers to the status parameters of the physical processing equipment during operation, including spindle load data, tool wear data, vibration data, and temperature data, reflecting the real-time operating status of the equipment. Status variables are digital parameters in the digital twin that correspond one-to-one with the physical equipment status data; they are the core carrier for achieving state synchronization of the digital twin, and the status variables are updated synchronously when the physical equipment status changes.

[0110] Specifically, firstly, corresponding monitoring sensors are deployed in key parts of the physical machining equipment: load sensors, vibration sensors, and temperature sensors are deployed in the spindle area, and wear monitoring sensors are deployed in the tool area. Then, an industrial communication protocol compatible with the physical equipment is selected to establish a real-time data transmission link between the sensors, the CNC system of the physical equipment, and the digital twin. Through this link, real-time equipment status data such as spindle load rate, tool wear VB value, spindle vibration amplitude, and spindle operating temperature are collected. Finally, the collected real-time data is accurately mapped to the corresponding state variables of the geometric model and physical model in the digital twin, realizing real-time synchronous updates of the state variables of the digital twin and the state data of the physical equipment.

[0111] For example, load sensors, vibration sensors, and temperature sensors are deployed at the spindle end of the aforementioned 3-axis CNC milling machine, and tool wear monitoring sensors are deployed at the tool clamping mechanism. The Profinet industrial communication protocol is selected to establish a real-time data transmission link between the sensors, the FANUC0i-MF CNC system of the machine, and the digital twin. Through this link, the spindle load rate, the wear VB value of the φ8mm carbide twist drill, the spindle vibration amplitude, and the spindle operating temperature are collected in real time. These real-time data are mapped to the spindle load state variables, tool wear state variables, spindle vibration state variables, and spindle temperature state variables in the digital twin, respectively. When the spindle load rate of the physical machine rises to 80% and the tool VB value increases to 0.025mm, the corresponding state variables of the digital twin are updated synchronously, achieving consistency with the state of the physical machine.

[0112] In this embodiment of the invention, a comprehensive and accurate mapping of physical machining equipment is achieved by constructing a three-dimensional digital twin integrating geometry, physics, and behavior: the geometric model restores the physical structure and motion relationships of the equipment, providing a realistic equipment shape carrier for virtual machining; the physical model incorporates machine tool dynamic characteristic parameters and tool wear evolution laws, allowing virtual machining to realistically reflect the equipment response under cutting forces and the impact of tool wear on machining, solving the problem of traditional simulation lacking physical characteristic support; the behavioral model replicates the interpolation logic and acceleration / deceleration control strategy of the CNC controller, ensuring a high degree of consistency between the motion trajectory and time axis of the virtual machining and the execution process of the physical equipment. Simultaneously, real-time acquisition of equipment status data and synchronous updating of digital twin state variables are achieved through industrial communication protocols, ensuring that the operating state of the digital twin remains dynamically consistent with the physical equipment. The final constructed digital twin provides a precise, real-time, and comprehensive virtual equipment carrier for subsequent virtual machining simulation, guaranteeing the realism of the virtual machining simulation process.

[0113] S500: Input the theoretical toolpath data into the digital twin, perform virtual machining simulation, obtain the simulation time, and calculate the deviation between the theoretical time and the simulation time.

[0114] In this embodiment of the invention, the theoretical toolpath data is input into the digital twin, virtual machining simulation is performed, simulation time is obtained, and the deviation between the theoretical time and the simulation time is calculated. The theoretical time is calculated based on standardized cutting parameters and static preset auxiliary time, where the preset auxiliary time is a fixed average value of historical machining statistics and is not coupled with the real-time operating status of the equipment. The auxiliary time in the simulation time is dynamically simulated through a behavioral model and may vary depending on the equipment status. For example, a high spindle temperature may cause the tool changer to move slowly, resulting in an inherent deviation between the simulation time and the theoretical time. The digital twin has achieved full-dimensional accurate mapping of the physical machining equipment and can synchronize the equipment status in real time. Inputting the theoretical toolpath data into it to perform virtual machining simulation can restore the machining process that closely resembles the actual operating status of the physical equipment in the virtual environment, resulting in a more realistic simulation time. By calculating the deviation between the theoretical time and the simulation time, the difference between the ideal calculated value and the virtual actual value can be quantified, providing a precise and quantitative basis for subsequent dynamic adjustment of the time calibration coefficient based on reinforcement learning. This is a key intermediate step in achieving accurate time prediction.

[0115] Virtual machining simulation refers to converting theoretical toolpath data into motion commands that a digital twin can recognize. This data is then input into a digital twin that has completed physical equipment mapping and state synchronization. The simulation utilizes the twin's geometric, physical, and behavioral models to fully reproduce the entire machining process of the physical equipment in a virtual environment. The simulation accurately reflects the actual impact of equipment dynamics, tool wear, and real-time status on the machining process. Simulation time refers to the actual time taken to complete the entire machining process of a target part using virtual machining simulation. This time incorporates real machining influencing factors such as the actual equipment state, dynamics, and tool wear, making it closer to the actual machining time of the physical equipment. Deviation value is a numerical value used to quantify the difference between theoretical and simulated machining time. It includes absolute and relative deviation values. The absolute deviation value is the absolute value of the difference between the theoretical and simulated machining time, reflecting the actual difference in time consumption. The relative deviation value is the ratio of the absolute deviation value to the theoretical machining time multiplied by 100%, reflecting the relative degree of deviation and serving as the quantitative basis for subsequent machining time calibration.

[0116] Specifically, the theoretical toolpath data generated in S300 is converted into a motion command format compatible with the digital twin of the target equipment. This allows the digital twin to recognize and execute the toolpath motion logic. During the conversion, core parameters such as toolpath length and feed rate for each process are preserved to ensure consistency in toolpath motion. The converted motion commands are then input into the digital twin of the target equipment to initiate the virtual machining simulation process. During the virtual machining simulation, the entire process time from the start of machining to the completion of all machining processes and the equipment returning to its initial machining state is recorded in real time. This time is the simulation time. The recording process accurately calculates the actual cutting time and the actual time consumed by various auxiliary operations, maintaining consistency with the actual machining time statistics logic. Based on the theoretical time obtained from S300 and the simulation time obtained in this simulation, the absolute deviation value is calculated: Absolute deviation value = |Theoretical time - Simulation time|. Then, the relative deviation value is calculated: Relative deviation value = (Absolute deviation value / Theoretical time) × 100%.

[0117] For example, using a digital twin of a 3-axis CNC milling machine for machining mold steel cores as a simulation platform, the theoretical machining time for the core in S300 is 2.823 min, and the theoretical toolpath data is as follows: drilling process toolpath 10 mm, feed 60 mm / min; chamfering process toolpath 20 mm, feed 200 mm / min; milling cavity process toolpath 100 mm, feed 180 mm / min. The above theoretical toolpath data is converted into a FANUCG code motion instruction format compatible with the digital twin of the 3-axis CNC milling machine, retaining the toolpath length and feed rate parameters for each process, and generating a series of motion instructions for drilling (G81), chamfering (G01), and milling (G41 / G42), ensuring that the digital twin can accurately identify the toolpath logic. The converted FANUCG code was input into the digital twin of the equipment, and the virtual machining simulation was started. During the simulation, the geometric model, physical model, and behavioral model of the equipment were called, and combined with the equipment status data synchronized in real time via the Profinet protocol, the cutting motion of the tool along the toolpath was reproduced. At the same time, the slight decrease in cutting efficiency caused by tool wear, the slight adjustment of feed rate due to spindle load limitation, and the acceleration and deceleration control timing of the equipment were also reproduced. The total time of the entire virtual machining simulation was recorded in real time. The actual cutting time of the drilling process was 0.167 min, the chamfering process was 0.11 min, the milling cavity process was 0.62 min, and the actual time of various auxiliary operations due to equipment status was 2.145 min. The total simulation time was 3.097 min, which is the simulation time of this virtual machining simulation. Based on the theoretical working time of 2.723 min and the simulated working time of 2.95 min, the absolute deviation value is calculated as |2.723-2.95|=0.227 min, and the relative deviation value is calculated as (0.227 / 2.723)×100%≈8.34%.

[0118] In this embodiment of the invention, by adapting and converting theoretical toolpath data, the digital twin accurately identifies the toolpath motion logic. The converted instructions are input into the digital twin, and simulation is initiated. Combined with real-time equipment status data, a three-in-one model is invoked to reproduce the processing process closely resembling the actual operating state of the physical equipment. This overcomes the limitations of traditional theoretical time calculations based solely on ideal conditions, resulting in simulated time that more closely matches the actual processing time, thus enhancing the practical reference value of the time data. Simultaneously, by quantifying the absolute and relative deviations between theoretical and simulated time, the degree of difference between the ideal calculated value and the virtual actual value is reflected. This provides a specific and quantifiable basis for the subsequent dynamic adjustment of time calibration coefficients by the reinforcement learning model, making subsequent time calibration more targeted and scientific. Furthermore, the virtual processing simulation is completed in a digital environment, eliminating the need for actual trial cutting with physical equipment. This effectively saves material and time costs associated with trial cutting, while avoiding equipment damage and processing defects during the trial cutting process. This improves the accuracy of time prediction while optimizing processing costs and efficiency.

[0119] S600: Based on the equipment status data, the processing feature information, and the deviation value, the working time calibration coefficient is dynamically adjusted through a reinforcement learning model to obtain the final estimated working time.

[0120] In this embodiment of the invention, based on the equipment status data, the processing feature information, and the deviation value, a reinforcement learning model is used to dynamically adjust the time calibration coefficient to obtain the final estimated time. The time calibration coefficient is the bridge connecting theoretical time and actual physical processing time. Traditional fixed calibration coefficients are based solely on historical average data statistics and cannot adapt to the real-time impact of dynamic equipment status and processing feature differences on time during a single batch of processing, resulting in the difficulty in continuously reducing the deviation between the final estimated time and the actual processing time. Reinforcement learning models, however, can dynamically adapt to time calibration requirements under different states through interactive learning with the environment. Therefore, it is necessary to construct a reinforcement learning model with equipment status, processing features, and deviation value as core inputs. Offline pre-training allows the model to master the decision-making logic of time calibration. Then, combined with real-time equipment status, processing features, and deviation value, the time calibration coefficient is dynamically adjusted to achieve accurate and adaptive optimization of the final estimated time, solving the problems of lag and inaccuracy in traditional fixed coefficient calibration.

[0121] Step S600 in the method provided in this embodiment of the invention includes:

[0122] Specifically, based on the equipment status data, the processing feature information, and the deviation value, the time calibration coefficient is dynamically adjusted through a reinforcement learning model to obtain the final estimated working hours, which includes the following steps:

[0123] Collect several historical processing task records for preset equipment models, extract historical equipment status record data, historical processing feature record data, historical theoretical working time record data, historical actual working time record data, and manually set historical calibration coefficient record data corresponding to each group of tasks, and form a historical experience sample set;

[0124] A reinforcement learning model is constructed, which includes an actor network and a critic network. The actor network is used to output the adjustment amount of the time calibration coefficient according to the current state, and the critic network is used to evaluate the value of the state-action pair.

[0125] The state space of the reinforcement learning model is set, wherein the state space includes a device state vector and a machining feature vector. The device state vector includes the spindle load mean, tool wear index, spindle temperature and vibration amplitude. The machining feature vector includes feature type encoding, material hardness, machining allowance and surface roughness requirements.

[0126] The action space of the reinforcement learning model is set, wherein the action space is the adjustment amount of the time calibration coefficient;

[0127] The reward function of the reinforcement learning model is set based on the final estimated working time and the actual processing time.

[0128] Based on the historical experience sample set, the reinforcement learning model is pre-trained offline using a deep deterministic policy gradient algorithm to obtain the reinforcement learning model.

[0129] The reward function of the reinforcement learning model is set as follows:

[0130] The reward function is calculated as follows: -|Final estimated working hours - Actual processing working hours|, where the final estimated working hours = theoretical working hours × (basic calibration coefficient + working hour calibration coefficient adjustment amount), and the basic calibration coefficient is obtained based on historical processing data.

[0131] The basic calibration coefficients are obtained based on historical processing data and include:

[0132] Obtain the theoretical and actual processing times of several historical parts of the same model as the target equipment, and calculate the ratio of the actual processing time to the theoretical processing time for each historical part; calculate the arithmetic mean of the ratios of several historical parts, and set the arithmetic mean as the basic calibration coefficient.

[0133] First, collect several historical processing task records for preset equipment models, and extract historical equipment status record data, historical processing feature record data, historical theoretical working time record data, historical actual working time record data, and manually set historical calibration coefficient record data corresponding to each group of tasks to form a historical experience sample set.

[0134] Among them, the preset equipment model refers to the equipment category that matches the target processing equipment model. Equipment of the same model has consistent dynamic characteristics and motion control logic, serving as a matching condition for historical sample collection. Historical processing task records refer to complete processing data records collected through CNC system interfaces, sensor monitoring, etc., after the target equipment completed historical processing tasks. Historical equipment status record data refers to real-time status data such as spindle load, tool wear, spindle temperature, and vibration amplitude collected by sensors during historical processing tasks. Historical processing feature record data refers to feature data such as fine-grained processing feature types, dimensions, positions, and material properties of historically processed parts. Historical theoretical processing time record data refers to the theoretical processing time data of historical parts calculated based on S300. Historical actual processing time record data refers to the actual processing time of historical parts collected through the CNC system interface, i.e., the difference between the processing start time and end time. Historical calibration coefficient record data refers to the manually set time calibration coefficients during historical processing, serving as the initial reference for offline pre-training of the reinforcement learning model. The historical experience sample set refers to the sample set formed by integrating the above-mentioned multi-dimensional historical data according to the task dimensions. Each sample corresponds to a complete processing task data, providing training support for the reinforcement learning model.

[0135] Specifically, first, the preset equipment model corresponding to the target equipment is determined, and 5,000 historical processing task records completed by this model of equipment in the past 3-5 years are collected. For each historical processing task, the corresponding data is extracted in the following way: the equipment status record data in the historical processing is collected through the industrial communication protocol, the historical theoretical working hours calculated by S300 are retrieved, the processing start and end times recorded by the CNC controller are automatically collected through the CNC system interface, the historical actual processing time is calculated, and the calibration coefficients manually set in the historical processing are sorted out. Finally, the equipment status data, processing feature data, theoretical working hours, actual working hours and calibration coefficients corresponding to each task are integrated one by one, and divided into training sample set, verification sample set and test sample set in a 7:2:1 ratio to form a complete historical experience sample set.

[0136] For example, taking a 3-axis CNC milling machine for machining mold steel cores as an example, 5000 historical machining task records of this machine were collected. For one historical machining task, the following data was extracted: equipment status record data: average spindle load 85 N·m, tool wear index 0.15, spindle temperature 75℃, vibration amplitude 0.045 mm; machining feature record data: blind hole feature φ8 mm, chamfer feature 45° φ5 mm, rectangular cavity feature 30 mm × 20 mm × 10 mm, material hardness HRC48; historical theoretical machining time 2.8 min, historical actual machining time 12.6 min, and manually set historical calibration coefficient 4.5. The above data were integrated into one sample. The same method was used to collect and integrate 5000 samples, which were divided into 3500 training samples, 1000 validation samples, and 500 test samples, forming a historical experience sample set.

[0137] Secondly, a reinforcement learning model is constructed, comprising an actor network and a critic network. The actor network outputs an adjustment amount for the time calibration coefficient based on the current state, while the critic network evaluates the value of the state-action pair. The reinforcement learning model is an intelligent model that achieves dynamic decision optimization through interactive learning with the environment. It consists of an actor network and a critic network. The actor network is responsible for selecting and executing actions in the current state, while the critic network evaluates the value of the selected actions based on reward signals from the environment. The two networks update their parameters through gradient information, guiding the decision strategy towards higher rewards in a continuous action-evaluation loop. The time calibration coefficient adjustment amount refers to the value used to fine-tune the basic calibration coefficient, ranging from -0.5 to 0.5. This adjustment amount dynamically corrects the basic calibration coefficient, adapting to the time calibration requirements of different processing scenarios.

[0138] For example, a deep neural network is used to construct the network structure of the reinforcement learning model: Actor network: The input layer is an 8-dimensional state vector, with two hidden layers. The first hidden layer contains 64 neurons, and the second hidden layer contains 32 neurons. The activation function is ReLU. The output layer has one neuron, with the output range limited to [-0.5, 0.5], corresponding to the adjustment amount of the work time calibration coefficient, ensuring a reasonable range of adjustment values. Critic network: The input layer contains an 8-dimensional state vector and a 1-dimensional action vector, which is the adjustment amount output by the actor network. It has two hidden layers, with the first hidden layer containing 64 neurons and the second hidden layer containing 32 neurons. The activation function is ReLU. The output layer has one neuron, outputting the value score of the state-action pair. A higher score indicates a better calibration effect of the adjustment amount. Finally, a reinforcement learning model containing actor and critic networks is constructed, ensuring that the actor can output accurate adjustment amounts, and the critic can effectively evaluate the adjustment amounts.

[0139] Next, the state space of the reinforcement learning model is defined, comprising a device state vector and a machining feature vector. The device state vector includes the average spindle load, tool wear index, spindle temperature, and vibration amplitude. The machining feature vector includes feature type encoding, material hardness, machining allowance, and surface roughness requirements. The state space refers to the set of all environmental state information that the reinforcement learning model can perceive when making decisions. The device state vector is a structured encoding vector of the real-time state data of the physical machining equipment, containing four dimensions: average spindle load, tool wear index, spindle temperature, and vibration amplitude. The machining feature vector is a structured encoding vector of the machining feature information of the target part, containing four dimensions: feature type encoding, material hardness, machining allowance, and surface roughness requirements.

[0140] Specifically, equipment status data and machining feature data are standardized and structurally integrated to construct an 8-dimensional feature vector as the state space of the reinforcement learning model. The equipment status vector and machining feature vector are each 4-dimensional. All continuous data are mapped to the 0-1 range using Min-Max normalization. The 4-dimensional equipment status vector normalizes the spindle load mean to the rated load range of the equipment, the tool wear index to the VB value threshold range of commonly used tools, the spindle temperature to the normal operating temperature range of the equipment, and the vibration amplitude to the allowable vibration amplitude threshold range of the equipment spindle. The 4-dimensional machining feature vector maps 9 types of fine-grained machining features to integer codes from 0 to 8. Material hardness is quantified by HB value, machining allowance by conventional machining allowance, and surface roughness requirement by Ra value, all normalized to the commonly used machining range of the equipment. The two types of vectors are sequentially concatenated to form an 8-dimensional fused feature vector, which serves as the input to the state space.

[0141] For example, taking the machining data of mold steel core as an example, the equipment state vector is: spindle load average 0.75, tool wear index 0.02, spindle temperature 0.65, vibration amplitude 0.03; the machining feature vector is: feature type code 2, material hardness 0.52, machining allowance 0.3, surface roughness 0.2; fused into an 8-dimensional state vector [0.75, 0.02, 0.65, 0.03, 2, 0.52, 0.3, 0.2], as the input of the state space.

[0142] Furthermore, the action space of the reinforcement learning model is set, wherein the action space is the adjustment amount of the time calibration coefficient. The action space refers to the set of all possible actions that the reinforcement learning model can output; in this embodiment, it is the set of values ​​for the time calibration coefficient adjustment amount. The action space is determined to be the value range of the time calibration coefficient adjustment amount [-0.5, 0.5], with a step size of 0.01. That is, the adjustment amount that the reinforcement learning model can output is -0.5, -0.49, ..., 0, ..., 0.49, 0.5, ensuring the refinement and rationality of the adjustment amount.

[0143] Simultaneously, a reward function for the reinforcement learning model is set, which is based on the final estimated working hours and the actual processing working hours.

[0144] The reward function of the reinforcement learning model is set as follows:

[0145] The reward function is calculated as follows: -|Final estimated working hours - Actual processing working hours|, where the final estimated working hours = theoretical working hours × (basic calibration coefficient + working hour calibration coefficient adjustment amount), and the basic calibration coefficient is obtained based on historical processing data.

[0146] The basic calibration coefficients are obtained based on historical processing data and include:

[0147] Obtain the theoretical and actual processing times of several historical parts of the same model as the target equipment, and calculate the ratio of the actual processing time to the theoretical processing time for each historical part; calculate the arithmetic mean of the ratios of several historical parts, and set the arithmetic mean as the basic calibration coefficient.

[0148] The reward function is an evaluation function used to quantify the effectiveness of the model's decision-making. It is set based on the difference between the final estimated working hours and the actual processing working hours. A higher reward value indicates a more accurate calibration decision by the model. The basic calibration coefficient is an initial calibration coefficient obtained based on statistical analysis of historical processing data of the target equipment. It is calculated as the average ratio of theoretical working hours to actual working hours in historical tasks and stored in the process knowledge base as the initial basic calibration coefficient.

[0149] First, calculate the basic calibration coefficient: obtain the theoretical and actual processing times of several historical parts of the same model as the target equipment, calculate the ratio of the actual processing time to the theoretical processing time for each historical part, and take the arithmetic mean of all ratios. Set the arithmetic mean as the basic calibration coefficient.

[0150] Then, define the reward function: Reward function = -|Final estimated working hours - Actual processing working hours|, where the final estimated working hours = theoretical working hours × (basic calibration coefficient + working hour calibration coefficient adjustment amount). The smaller the absolute value of the reward value, the closer the final estimated working hours are to the actual processing working hours, and the better the model decision-making effect.

[0151] For example, 1000 historical parts of the target device are collected, and the ratio of actual working time to theoretical working time for each part is calculated. The arithmetic mean yields a basic calibration coefficient of 4.3. If the output adjustment of the reinforcement learning model is 0.12, the theoretical working time is known to be 2.723 min, and the actual working time is 2.95 min, then the final estimated working time = 2.723 × (1.03 + 0.12) ≈ 3.131 min, the reward function = -|3.131 - 2.95| = ​​-0.181, and the reward value is -0.181.

[0152] Subsequently, based on the historical experience sample set, the reinforcement learning model is pre-trained offline using the Deep Deterministic Policy Gradient (DDPG) algorithm to obtain the reinforcement learning model. The Deep Deterministic Policy Gradient (DDPG) algorithm is a reinforcement learning algorithm suitable for continuous action spaces. It outputs deterministic actions through an actor network, evaluates the value of actions using a critic network, and utilizes experience replay and target network update mechanisms to achieve efficient offline training of the model, allowing the model to gradually learn the optimal decision-making policy.

[0153] Specifically, based on a historical experience sample set, the DDPG algorithm is used to pre-train the reinforcement learning model offline. The training process is as follows: Initialize the actor network, critic network, and corresponding target actor network and target critic network, with the target network parameters being consistent with the initial parameters of the original network; initialize the experience replay pool with a capacity of 10,000 samples; randomly select samples with a batch size of 64 from the training samples in the historical experience sample set, input the state vectors from the samples into the actor network to obtain the adjustment amount of the work time calibration coefficient; input the state vectors and adjustment amounts into the critic network to obtain the action value score; calculate the reward function value and combine it with the target network to calculate the target value, and update the parameters of the actor and critic networks through backpropagation; after every 100 training steps, softly update the parameters of the original actor and critic networks to the target network; repeat the above steps until the average reward value of the reinforcement learning model on the validation sample set does not improve for 50 consecutive rounds, or reaches a preset number of training rounds, such as 1000 rounds, to complete the offline pre-training of the model and obtain a reinforcement learning model that can be used for online decision-making.

[0154] For example, based on the historical experience sample set of a 3-axis CNC milling machine, the DDPG algorithm is used for offline pre-training: the actor, critic, and target networks are initialized, and the experience replay pool capacity is set to 10,000; 64 sets of samples are randomly selected each time, input into the actor network to obtain the adjustment amount, input into the critic network to obtain the value score, and the target value is calculated by combining the reward function value, and the network parameters are updated in reverse; the target network is softly updated every 100 steps. When training reaches the 650th round, the average reward value of the validation set stabilizes from the initial -0.5 to -0.08, and there is no improvement for 50 consecutive rounds. Training is stopped, and the pre-training of the reinforcement learning model is completed.

[0155] Finally, based on the equipment status data, the processing feature information, and the deviation value, the time calibration coefficient is dynamically adjusted through a reinforcement learning model to obtain the final estimated time. Online decision-making refers to the process of inputting real-time equipment status data and processing feature information of the target part processing into a pre-trained reinforcement learning model, and having the actor network output the adjustment amount of the time calibration coefficient in real time. The final estimated time refers to the accurate time prediction result calculated by combining the theoretical time, the basic calibration coefficient, and the adjustment amount output by the actor network; it is the final time data that integrates ideal calculations, historical experience, and real-time status.

[0156] Specifically, real-time equipment status data and processing feature information are collected and converted into an 8-dimensional state vector, which is then input into the pre-trained reinforcement learning model. The actor network outputs the adjustment amount of the time calibration coefficient within the range of [-0.5, 0.5] based on the state vector. The basic calibration coefficients stored in the process knowledge base are retrieved to calculate the final estimated time of the target part. The final estimated time of this processing is recorded. After processing is completed, the actual processing time is collected through the CNC system interface for subsequent online iterative optimization of the model.

[0157] For example, real-time data on the equipment status during core machining of mold steel is collected: spindle load average 0.9, tool wear index 0.18, spindle temperature 0.85, vibration amplitude 0.06; machining feature information: blind hole features, material hardness HRC48, machining allowance 0.3, surface roughness 0.2; an 8-dimensional state vector [0.8, 0.025, 0.7, 0.035, 2, 0.52, 0.3, 0.2] is constructed and input into a pre-trained reinforcement learning model; the actor network outputs a time calibration coefficient adjustment of 0.2; the basic calibration coefficient of 4.3 is retrieved, the theoretical time is 2.823 min, and the final estimated time is calculated as 2.823 × (4.3 + 0.2) ≈ 12.7 min.

[0158] In this embodiment of the invention, a historical experience sample set is constructed by collecting a large number of historical processing task records, providing sufficient and comprehensive training data support for the reinforcement learning model. This ensures that the model can cover decision-making scenarios with different processing characteristics and equipment states. The constructed actor-critic dual-network reinforcement learning model realizes the closed-loop logic of decision-making and evaluation, allowing the reinforcement learning model to accurately output the adjustment amount of the time calibration coefficient adapted to different scenarios. By setting the state space, action space, and reward function, combined with the offline pre-training of the DDPG algorithm, the reinforcement learning model can quickly learn the optimal time calibration decision strategy, effectively improving the model's decision stability and accuracy. Finally, based on the real-time equipment state, processing characteristics, and pre-trained model, the time calibration coefficient is dynamically adjusted, breaking through the limitations of the traditional fixed calibration coefficient, realizing adaptive and dynamic optimization of time estimation, and ensuring that the final estimated time accurately matches the actual processing time of the physical equipment.

[0159] The method provided in this embodiment of the invention further includes:

[0160] When the deviation value is greater than a preset deviation threshold, deviation cause diagnosis is triggered, wherein the deviation cause diagnosis includes identifying the reduction in cutting speed due to tool wear and the adjustment of feed rate due to spindle load limitation;

[0161] The diagnostic results of the deviation are added to the device status data for subsequent reinforcement learning model training.

[0162] First, when the deviation value exceeds a preset deviation threshold, deviation cause diagnosis is triggered. This deviation cause diagnosis includes identifying reductions in cutting speed due to tool wear and feed rate adjustments due to spindle load limitations. The relative deviation between the theoretical and simulated machining times calculated in S500 is compared with the preset deviation threshold. If the relative deviation value exceeds the preset deviation threshold, the deviation cause diagnosis process is immediately triggered. By analyzing the actual values ​​of the tool wear VB value and the average spindle load in the real-time equipment status data, combined with the differences between the actual and theoretical values ​​of cutting speed and feed rate during the virtual machining simulation, specific cause identification is completed: if the tool wear VB value exceeds the normal operating threshold, and the actual cutting speed in the simulation is lower than the theoretical value, it is determined that the reduction in cutting speed is due to tool wear; if the average spindle load continuously exceeds the equipment's rated load threshold, and the actual feed rate in the simulation is lower than the theoretical value, it is determined that the feed rate adjustment is due to spindle load limitations.

[0163] For example, with a preset deviation threshold of 8%, S500 calculates a relative deviation of 8.34% for machining the mold steel core, which is greater than the preset deviation threshold, triggering deviation cause diagnosis. Analyzing real-time equipment status data, the tool wear VB value is 0.18mm, and the normal operating threshold is 0.02mm; the average spindle load is 80%, and the equipment rated load threshold is 75%. Combining virtual machining simulation data, the actual cutting speed of 900r / min is lower than the theoretical value of 1000r / min, and the actual feed rate of 50mm / min is lower than the theoretical value of 60mm / min. Based on this, the specific cause of this deviation is identified as: reduced cutting speed due to tool wear and feed rate adjustment due to spindle load limitation.

[0164] Secondly, the diagnostic results of the deviation causes are added to the equipment status data for subsequent reinforcement learning model training. The specific causes of the deviations and the corresponding parameter exceedances and processing parameter reductions are structured and encoded. The encoded diagnostic results are added to the original equipment status dataset to form expanded equipment status data. The expanded equipment status data is then integrated with corresponding processing feature information, deviation values, actual processing time, and other data to form new training samples, which are included in the historical experience sample set for subsequent online iterative training and optimization of the reinforcement learning model.

[0165] For example, the structured encoding of this deviation diagnosis result is as follows: tool wear causes a decrease in cutting speed, VB value exceeds the threshold by 0.005mm, and the cutting speed decreases by 6.67%; spindle load limitation causes feed rate adjustment, load exceeds the threshold by 5%, and the feed rate decreases by 6.67%. This encoded information is added to the original equipment status data, which includes spindle load average 0.8, tool wear index 0.025, spindle temperature 0.7, and vibration amplitude 0.035, to form expanded equipment status data. This expanded data is then integrated with the machining characteristics of the mold steel core, the relative deviation value of 8.34%, and the actual machining time of 3.01min to form a new training sample, which is included in the historical experience sample set for subsequent online iterative training of the reinforcement learning model.

[0166] In this embodiment of the invention, a preset deviation threshold is used to accurately identify and diagnose significant deviation scenarios, quickly pinpointing the specific equipment cause of time deviations. This solves the problems of vague deviation causes and lack of targeted optimization in traditional time estimation. Integrating structured diagnostic results into equipment status data effectively enriches the dimensions of this data, ensuring it not only includes basic operating parameters but also causal information linking parameter anomalies to changes in processing parameters. This provides more targeted abnormal scenario samples for reinforcement learning model training. Simultaneously, incorporating new training samples into the historical experience sample set drives online iterative optimization of the reinforcement learning model. This allows the model to gradually learn the correlation between abnormal equipment states and time deviations, improving its ability to calibrate time for abnormal processing scenarios. This further reduces the deviation between the final estimated time and the actual processing time, continuously optimizing the adaptability and accuracy of the entire time estimation method.

[0167] Through the specific implementation methods described above, the embodiments of the present invention achieve the following technical effects:

[0168] This invention provides a data analysis-driven intelligent prediction method and system for machining time. By loading 3D model data of the target part, identifying multimodal fine-grained features, making process decisions and calculating theoretical machining time, constructing and loading a digital twin of the equipment in real-time synchronous status, obtaining simulated machining time through virtual machining simulation and calculating deviation values, and dynamically adjusting the time calibration coefficient through reinforcement learning, the entire machining time prediction process is automated and intelligent. It provides high-quality, two-dimensional basic data for subsequent stages from the source, accurately extracting fine-grained machining features that fit the actual process, making process planning and theoretical time calculation more scientific and standardized. The digital twin recreates the real machining scenario and quantifies time deviations. Finally, dynamic calibration replaces traditional static coefficients, improving the accuracy of time prediction, effectively reducing the deviation between predicted and actual machining times, and improving time prediction efficiency. This provides accurate and efficient data support for enterprise cost accounting, production scheduling, and pricing decisions.

[0169] Example 2, as Figure 3 , Figure 4 As shown, this invention provides a data analysis-driven intelligent prediction system for machining time, the system comprising:

[0170] The 3D model loading module 11 is used to load the 3D model data of the target part, wherein the 3D model data includes geometric topology information and surface visual information;

[0171] The multimodal feature recognition module 12 is used to perform multimodal feature recognition based on the geometric topology information and the surface visual information to obtain processing feature information, wherein the processing feature information includes feature type, feature size and feature position;

[0172] The process decision and time calculation module 13 is used to perform process decision and theoretical time calculation based on the processing feature information to obtain theoretical time and theoretical toolpath data;

[0173] The digital twin loading module 14 is used to load a digital twin of the target device. The digital twin includes a geometric model, a physical model, and a behavioral model, and synchronizes the device status data with the physical processing device in real time. The device status data includes spindle load data, tool wear data, vibration data, and temperature data.

[0174] The virtual machining simulation module 15 is used to input the theoretical toolpath data into the digital twin, perform virtual machining simulation, obtain the simulation time, and calculate the deviation value between the theoretical time and the simulation time.

[0175] The intelligent time calibration module 16 is used to dynamically adjust the time calibration coefficient based on the equipment status data, the processing feature information and the deviation value through a reinforcement learning model to obtain the final estimated time.

[0176] In one embodiment, the 3D model loading module 11 is further configured to:

[0177] Obtain the solid 3D model file of the target part;

[0178] Perform geometric integrity verification on the entity 3D model file to obtain a verified entity 3D model.

[0179] The verified 3D model of the entity is parsed, and the geometric topology information is extracted, wherein the geometric topology information is represented in the form of an attribute adjacency graph;

[0180] The verified 3D entity model is rendered in 3D voxelization to generate voxel mesh data, which is then added to the surface visual information.

[0181] In one embodiment, the multimodal feature recognition module 12 is further configured to:

[0182] Collect several historically processed parts of a preset part model. For each historically processed part, extract its geometric topology record data and surface visual record data. At the same time, label the first fine-grained feature identifier up to the Pth fine-grained feature identifier contained in the historically processed part, where P is the total number of preset fine-grained processing feature categories. The fine-grained processing features include through hole features, blind hole features, countersunk hole features, U-groove features, T-groove features, dovetail groove features, rectangular cavity features, circular cavity features, and chamfer features.

[0183] Using the first fine-grained feature identifier as supervision and the geometric topology recording data and the surface visual recording data as input, a first feature recognizer capable of recognizing the first fine-grained feature is trained.

[0184] The process continues until the Pth fine-grained feature identifier is used as supervision, and the geometric topology recording data and the surface visual recording data are used as inputs to train a Pth feature recognizer capable of recognizing the Pth fine-grained features.

[0185] Merge the output layers of the first feature recognizer up to the Pth feature recognizer to construct a multi-label classification model, obtain a multi-modal feature recognition model, store it in association with the preset part model, and add it to the multi-modal feature recognition model set;

[0186] Based on the target part model, the multimodal feature recognition model set is invoked, and the geometric topology information and surface visual information of the current part are input into the multimodal feature recognition model. Through forward propagation, the probability values ​​of the current part belonging to the first processing feature up to the Pth processing feature are obtained simultaneously. The processing features with a probability value greater than a preset probability threshold are taken as the recognition result, and the processing feature information is output.

[0187] In one embodiment, the process decision and time calculation module 13 is further configured to:

[0188] Load a process knowledge graph, wherein the process knowledge graph constructs a semantic network using part feature nodes, machining process nodes, tool nodes, and material nodes;

[0189] The processing feature information is input into the process knowledge graph, and similar historical parts are matched to obtain the initial process chain.

[0190] Based on the initial process chain and the preset cutting parameter library, theoretical toolpath data is generated, wherein the theoretical toolpath data includes the toolpath length and feed rate of each process;

[0191] Based on the theoretical toolpath data, the theoretical machining time is calculated and a preset auxiliary time is added to obtain the theoretical machining time.

[0192] In one embodiment, the digital twin loading module 14 is further configured to:

[0193] A geometric model is established for the physical processing equipment, wherein the geometric model includes the three-dimensional structure of the machine tool bed, spindle, tool magazine and worktable;

[0194] A physical model is established for the physical processing equipment, wherein the physical model includes dynamic characteristic parameters and tool wear evolution law;

[0195] A behavioral model is established for the physical processing equipment, wherein the behavioral model simulates the interpolation logic and acceleration / deceleration control of the CNC controller;

[0196] The device status data is collected in real time using industrial communication protocols, and the status variables of the digital twin are updated synchronously.

[0197] In one embodiment, the working time intelligent calibration module 16 is further configured to:

[0198] Collect several historical processing task records for preset equipment models, extract historical equipment status record data, historical processing feature record data, historical theoretical working time record data, historical actual working time record data, and manually set historical calibration coefficient record data corresponding to each group of tasks, and form a historical experience sample set;

[0199] A reinforcement learning model is constructed, which includes an actor network and a critic network. The actor network is used to output the adjustment amount of the time calibration coefficient according to the current state, and the critic network is used to evaluate the value of the state-action pair.

[0200] The state space of the reinforcement learning model is set, wherein the state space includes a device state vector and a machining feature vector. The device state vector includes the spindle load mean, tool wear index, spindle temperature and vibration amplitude. The machining feature vector includes feature type encoding, material hardness, machining allowance and surface roughness requirements.

[0201] The action space of the reinforcement learning model is set, wherein the action space is the adjustment amount of the time calibration coefficient;

[0202] The reward function of the reinforcement learning model is set based on the final estimated working time and the actual processing time.

[0203] Based on the historical experience sample set, the reinforcement learning model is pre-trained offline using a deep deterministic policy gradient algorithm to obtain the reinforcement learning model.

[0204] The reward function of the reinforcement learning model is set as follows:

[0205] The reward function is calculated as follows: -|Final estimated working hours - Actual processing working hours|, where the final estimated working hours = theoretical working hours × (basic calibration coefficient + working hour calibration coefficient adjustment amount), and the basic calibration coefficient is obtained based on historical processing data.

[0206] The basic calibration coefficients are obtained based on historical processing data and include:

[0207] Obtain the theoretical and actual processing times of several historical parts of the same model as the target equipment, and calculate the ratio of the actual processing time to the theoretical processing time for each historical part; calculate the arithmetic mean of the ratios of several historical parts, and set the arithmetic mean as the basic calibration coefficient.

[0208] The system provided in this embodiment of the invention is also used for:

[0209] When the deviation value is greater than a preset deviation threshold, deviation cause diagnosis is triggered, wherein the deviation cause diagnosis includes identifying the reduction in cutting speed due to tool wear and the adjustment of feed rate due to spindle load limitation;

[0210] The diagnostic results of the deviation are added to the device status data for subsequent reinforcement learning model training.

[0211] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A data analytics driven intelligent estimation of machining labor hours method, characterized in that, include: Load the three-dimensional model data of the target part, wherein the three-dimensional model data includes geometric topology information and surface visual information; Based on the geometric topology information and the surface visual information, multimodal feature recognition is performed to obtain processing feature information, wherein the processing feature information includes feature type, feature size and feature position; Based on the machining feature information, process decision-making and theoretical time calculation are performed to obtain theoretical time and theoretical toolpath data; A digital twin of the target device is loaded, wherein the digital twin includes a geometric model, a physical model and a behavioral model, and the device status data is synchronized with the physical processing device in real time, the device status data including spindle load data, tool wear data, vibration data and temperature data; The theoretical toolpath data is input into the digital twin, virtual machining simulation is performed, simulation time is obtained, and the deviation between the theoretical time and the simulation time is calculated. Based on the equipment status data, the processing feature information, and the deviation value, the time calibration coefficient is dynamically adjusted through a reinforcement learning model to obtain the final estimated working time.

2. The data analytics driven machining labor intelligence estimation method of claim 1, wherein, Load the 3D model data of the target part, including: Obtain the solid 3D model file of the target part; Perform geometric integrity verification on the entity 3D model file to obtain a verified entity 3D model. The verified 3D model of the entity is parsed, and the geometric topology information is extracted, wherein the geometric topology information is represented in the form of an attribute adjacency graph; The verified 3D entity model is rendered in 3D voxelization to generate voxel mesh data, which is then added to the surface visual information.

3. The data analytics driven machining labor intelligence estimation method of claim 1, wherein, Based on the geometric topology information and the surface visual information, multimodal feature recognition is performed to obtain processing feature information, including the following: Collect several historically processed parts of a preset part model. For each historically processed part, extract its geometric topology record data and surface visual record data. At the same time, label the first fine-grained feature identifier up to the Pth fine-grained feature identifier contained in the historically processed part, where P is the total number of preset fine-grained processing feature categories. The fine-grained processing features include through hole features, blind hole features, countersunk hole features, U-groove features, T-groove features, dovetail groove features, rectangular cavity features, circular cavity features, and chamfer features. Using the first fine-grained feature identifier as supervision and the geometric topology recording data and the surface visual recording data as input, a first feature recognizer capable of recognizing the first fine-grained feature is trained. The process continues until the Pth fine-grained feature identifier is used as supervision, and the geometric topology recording data and the surface visual recording data are used as inputs to train a Pth feature recognizer capable of recognizing the Pth fine-grained features. Merge the output layers of the first feature recognizer up to the Pth feature recognizer to construct a multi-label classification model, obtain a multi-modal feature recognition model, store it in association with the preset part model, and add it to the multi-modal feature recognition model set; Based on the target part model, the multimodal feature recognition model set is invoked, and the geometric topology information and surface visual information of the current part are input into the multimodal feature recognition model. Through forward propagation, the probability values ​​of the current part belonging to the first processing feature up to the Pth processing feature are obtained simultaneously. The processing features with a probability value greater than a preset probability threshold are taken as the recognition result, and the processing feature information is output.

4. The data analytics driven machining labor intelligence estimation method of claim 1, wherein, Based on the machining feature information, process decisions and theoretical time calculations are performed to obtain theoretical time and theoretical toolpath data, including: Load a process knowledge graph, wherein the process knowledge graph constructs a semantic network using part feature nodes, machining process nodes, tool nodes, and material nodes; The processing feature information is input into the process knowledge graph, and similar historical parts are matched to obtain the initial process chain. Based on the initial process chain and the preset cutting parameter library, theoretical toolpath data is generated, wherein the theoretical toolpath data includes the toolpath length and feed rate of each process; Based on the theoretical toolpath data, the theoretical machining time is calculated and a preset auxiliary time is added to obtain the theoretical machining time.

5. The data analytics driven machining labor intelligence estimation method of claim 1, wherein, Load the digital twin of the target device, including: A geometric model is established for the physical processing equipment, wherein the geometric model includes the three-dimensional structure of the machine tool bed, spindle, tool magazine and worktable; A physical model is established for the physical processing equipment, wherein the physical model includes dynamic characteristic parameters and tool wear evolution law; A behavioral model is established for the physical processing equipment, wherein the behavioral model simulates the interpolation logic and acceleration / deceleration control of the CNC controller; The device status data is collected in real time using industrial communication protocols, and the status variables of the digital twin are updated synchronously.

6. The data analytics driven machining labor intelligence estimation method of claim 1, wherein, Based on the equipment status data, the processing feature information, and the deviation value, the time calibration coefficient is dynamically adjusted using a reinforcement learning model to obtain the final estimated working time, which includes the following steps: Collect several historical processing task records for preset equipment models, extract historical equipment status record data, historical processing feature record data, historical theoretical working time record data, historical actual working time record data, and manually set historical calibration coefficient record data corresponding to each group of tasks, and form a historical experience sample set; A reinforcement learning model is constructed, which includes an actor network and a critic network. The actor network is used to output the adjustment amount of the time calibration coefficient according to the current state, and the critic network is used to evaluate the value of the state-action pair. The state space of the reinforcement learning model is set, wherein the state space includes a device state vector and a machining feature vector. The device state vector includes the spindle load mean, tool wear index, spindle temperature and vibration amplitude. The machining feature vector includes feature type encoding, material hardness, machining allowance and surface roughness requirements. The action space of the reinforcement learning model is set, wherein the action space is the adjustment amount of the time calibration coefficient; The reward function of the reinforcement learning model is set based on the final estimated working time and the actual processing time. Based on the historical experience sample set, the reinforcement learning model is pre-trained offline using a deep deterministic policy gradient algorithm to obtain the reinforcement learning model.

7. The data analysis-driven intelligent prediction method for machining time according to claim 6, characterized in that, Setting the reward function for the reinforcement learning model includes: The reward function is calculated as follows: -|Final estimated working hours - Actual processing working hours|, where the final estimated working hours = theoretical working hours × (basic calibration coefficient + working hour calibration coefficient adjustment amount), and the basic calibration coefficient is obtained based on historical processing data.

8. The data analysis-driven intelligent prediction method for machining time according to claim 7, characterized in that, The basic calibration coefficients are obtained based on historical processing data, including: Obtain the theoretical and actual processing times of several historical parts of the same model as the target equipment, and calculate the ratio of the actual processing time to the theoretical processing time for each historical part; calculate the arithmetic mean of the ratios of several historical parts, and set the arithmetic mean as the basic calibration coefficient.

9. The data analysis-driven intelligent prediction method for machining time as described in claim 1, characterized in that, Also includes: When the deviation value is greater than a preset deviation threshold, deviation cause diagnosis is triggered, wherein the deviation cause diagnosis includes identifying the reduction in cutting speed due to tool wear and the adjustment of feed rate due to spindle load limitation; The diagnostic results of the deviation are added to the device status data for subsequent reinforcement learning model training.

10. A data analysis-driven intelligent prediction system for machining time, characterized in that, The data analysis-driven intelligent prediction method for machining time according to any one of claims 1-9 includes: A 3D model loading module is used to load the 3D model data of the target part, wherein the 3D model data includes geometric topology information and surface visual information; A multimodal feature recognition module is used to perform multimodal feature recognition based on the geometric topology information and the surface visual information to obtain processing feature information, wherein the processing feature information includes feature type, feature size and feature position; The process decision and time calculation module is used to perform process decision and theoretical time calculation based on the machining feature information to obtain theoretical time and theoretical toolpath data; A digital twin loading module is used to load a digital twin of a target device. The digital twin includes a geometric model, a physical model, and a behavioral model, and synchronizes device status data with the physical processing device in real time. The device status data includes spindle load data, tool wear data, vibration data, and temperature data. The virtual machining simulation module is used to input the theoretical toolpath data into the digital twin, perform virtual machining simulation, obtain the simulation time, and calculate the deviation between the theoretical time and the simulation time. The intelligent time calibration module is used to dynamically adjust the time calibration coefficient based on the equipment status data, the processing feature information and the deviation value through a reinforcement learning model to obtain the final estimated working time.

Images