Construction of body model of precision numerical control machine tool and autonomous decision method and system and embodied intelligent system

Abstract

This invention discloses a method, system, and embodied intelligent system for constructing an embodied world model and making autonomous decisions for precision CNC machine tools. The method collects force, vibration, temperature, and multi-view images to construct a PA-3DGS twin field infused with CAD semantics, reconstructed by neural operators, and anchored to physical perception channels element by element. After latent state encoding, the method uses a forward dynamic world model to recursively predict future machining states based on candidate machining actions. It then performs imaginative autonomous decision-making planning on candidate action sequences and selects the best option based on multi-objective costs. A dual-timescale asynchronous dual-core architecture consisting of a cognitive kernel and an execution kernel completes imaginative planning and millisecond-level servo compensation under singular perturbation stability conditions. Finally, the method uses online continuous learning of prediction-measurement residuals to fine-tune the world model and reverse-anchor rendering, forming a self-consistent closed loop of perception-imagination-planning-execution-self-learning, enabling the precision machine tool to leap to an embodied intelligent agent capable of forward-looking deduction and autonomous decision-making.

Description

Technical Field

[0001] This invention belongs to the field of intelligent manufacturing and embodied intelligence technology. Specifically, it is a method, system and embodied intelligence system for constructing an embodied world model of a precision CNC machine tool and making autonomous decisions. It is particularly suitable for the forward prediction and autonomous decision-making control of the processing status of high-value parts such as aero-engine blades and precision gear pairs in the precision milling, grinding and gear processing process. Background Technology

[0002] High-value, complex parts, such as aero-engine blades and precision gear pairs, are core components in high-end manufacturing fields like aerospace and precision machinery. The precision machining of these parts typically relies on five-axis CNC machine tools, employing processes such as precision milling, grinding, and gear machining. However, these machining processes involve variable tool postures, low workpiece stiffness, large material removal volumes, and intense thermo-mechanical-vibration multi-physics coupling effects. These factors easily induce chatter and thermo-mechanical deformation, leading to surface quality degradation and geometric accuracy loss, and in severe cases, even workpiece scrap. How to proactively suppress and avoid errors during machining, before they irreversibly infiltrate the workpiece, is a critical technical challenge that urgently needs to be overcome in the high-end precision manufacturing field.

[0003] To address these challenges, the industry has developed a variety of technological approaches, mainly including the following three categories.

[0004] Category 1: Processing status monitoring technology based on digital twins / digital shadows.

[0005] Existing digital twin solutions for CNC machine tools mostly realize a one-way data flow from the physical entity to the virtual space in the form of a "digital shadow." The typical implementation involves mirroring motion and cutting based on a pre-established geometric model through rasterization or voxel Boolean operations, focusing on equipment status monitoring, machining process visualization, and predictive maintenance. The fundamental flaw of this approach is that it merely "reproduces" and "replays" the current and historical states, lacking the ability to predict how the machining state will evolve if a certain machining action is performed.

[0006] The second category is passive / reactive closed-loop control technology based on real-time sensor feedback.

[0007] Currently, the widely adopted machining quality control method in industry involves arranging force sensors, vibration sensors, and temperature sensors on machine tools to monitor physical signals during the cutting process in real time. When a signal exceeds a limit, an alarm, speed reduction, or machine shutdown is triggered. This method is a typical "sensor-response" passive regulation, and its core shortcomings are: First, the generation of control commands is based on "already occurred" error signals. The system can only respond after the error occurs and cannot prevent it before it forms. When the sensor detects abnormal cutting force or chatter signs, the error has often been partially or irreversibly etched into the workpiece surface. Second, existing reactive regulation is mostly based on rule-driven strategies such as threshold comparison, lookup tables, or fuzzy control, lacking the ability to "virtually test" the consequences of candidate process parameters, and thus failing to achieve truly forward-looking autonomous decision-making. Third, there is often a significant time mismatch between the inference delay of deep prediction models (such as neural networks) and the millisecond-level servo execution cycle, making it difficult to effectively convert prediction results into real-time control commands.

[0008] The third category: data-driven processing status prediction technology.

[0009] In recent years, with the rapid development of deep learning technology, data-driven machining state prediction methods have received widespread attention. These methods collect large amounts of historical machining data to train neural network models to predict tool wear, surface quality, or machining errors. However, existing data-driven methods have the following shortcomings: First, most methods model machining quality prediction as a single-step regression from the current signal to the current quality index, lacking temporal modeling of state evolution and multi-step look-ahead capabilities. Second, purely data-driven methods lack physical constraints; when the operating conditions exceed the distribution range of the training data, the prediction results may deviate significantly from the physical reality, making reliability difficult to guarantee. Third, data-driven models are mostly trained and deployed offline, lacking online continuous learning and self-evolution capabilities, making it difficult to adapt to gradual drift caused by tool wear and environmental changes during long-term operation. While existing solutions involve tool wear life prediction based on physical information neural networks, their predictions are based on degradation trend fitting rather than multi-step dynamic deduction conditioned on candidate machining actions. Although there are attempts at multi-task prediction (remaining life + anomaly probability), these are still single-step static predictions and have not constructed a forward deduction and imaginative planning mechanism conditioned on candidate actions.

[0010] In summary, existing technologies suffer from the following main shortcomings. First, a lack of foresight: Current digital twin technologies largely remain at the level of "mirror replication," only reflecting current and historical states, lacking the ability to predict how future processing states will evolve if a certain processing action is adopted. Second, passive control strategies: Existing closed-loop control is mostly reactive, only responding to errors that have already occurred, unable to prevent errors before they form, and lacking the ability to "virtually test and error" the consequences of candidate process parameters. Third, insufficient information fusion: Processing quality is affected not only by spatially distributed physical fields such as cutting force, vibration, and temperature, but also by global operating conditions such as ambient temperature and humidity, air pressure, and cutting fluid conditions, as well as slowly changing state factors such as tool wear and initial workpiece stress. Existing solutions either ignore these factors or treat them indiscriminately along with the spatial physical fields, leading to model redundancy and inconsistent predictions. Fourth, temporal mismatch: There is a significant temporal mismatch between the inference latency of deep prediction models and the millisecond-level servo execution cycle, making it difficult to effectively convert prediction results into real-time control commands. Fifth, lack of self-evolution capability: Most existing models are trained and deployed offline, lacking online continuous learning and self-correction mechanisms based on prediction-measurement residuals.

[0011] In view of this, how to deeply integrate an embodied world model with the physical perception twin of a precision machine tool, possessing the capabilities of "perception, imagination, planning, and self-evolution," while ensuring physical reliability and millisecond-level real-time performance, so that the precision machine tool itself truly becomes an embodied intelligent entity capable of forward-looking deduction and autonomous decision-making, is a pressing technical problem that needs to be solved. This invention proposes a method and system for constructing an embodied world model and enabling autonomous decision-making for precision CNC machine tools, precisely to address the aforementioned problem. Summary of the Invention

[0012] In view of this, the purpose of this invention is to provide a method, system, and embodied intelligent system for constructing an embodied world model and making autonomous decisions for precision CNC machine tools. By constructing a physically perceptual three-dimensional Gaussian sputtering twin field as a unified representation base, and combining a forward dynamic world model of motion conditions with imaginative autonomous decision-making planning, and with a dual-timescale asynchronous dual-core architecture, it realizes forward-looking deduction, autonomous decision-making, and closed-loop self-learning of the machining process. This fundamentally solves the problems of lack of forward-looking prediction, passive control response, and time mismatch in existing technologies, making precision machine tools into embodied intelligent entities capable of forward-looking deduction and autonomous decision-making.

[0013] To achieve the above objectives, the present invention provides the following technical solution: This invention first proposes a method for constructing an embodied world model and making autonomous decisions for a precision CNC machine tool, comprising the following steps: Step 1: Construction of Physically Perceived Embodied Representation: Multimodal sensors are deployed on a precision CNC machine tool and workpiece to collect cutting force, vibration, and temperature signals, as well as multi-view images during the machining process. Based on the multi-view images, the spatial geometry of the machine tool and workpiece is reconstructed using 3D Gaussian sputtering with an explicit radiation field. Computer-aided design (CAD) component-level semantic tags are then injected into the reconstructed Gaussian primitives. The cutting force, vibration, and temperature signals are reconstructed into a continuous physical field using neural operators. The physical field values ​​at the center of each Gaussian primitive are then anchored to the corresponding primitive, and these physical field values ​​are updated synchronously with the geometric pose changes of the Gaussian primitives. This results in a physically perceived 3D Gaussian sputtered twin field with augmented physical perception channels based on standard geometry and photometric properties, serving as the unified spatial anchoring representation base for the embodied world model. Step 2: Embodied Hidden State Encoding: The physically perceived 3D Gaussian sputtering twin field and multimodal signal are compressed into a compact hidden representation by an encoder, and then concatenated with the tool pose, current process parameters, global environment state vector representing the machining environment state, and slowly varying hidden state representing the slowly varying characteristics of the tool and workpiece to obtain the system state at the current moment. Step 3: Forward dynamics imagination of action conditions: Using the candidate processing action, the global environment state vector and the slowly changing hidden state as conditions, the learned forward dynamics world model is used to recursively imagine and predict the system state at the next moment and even the next H steps from the current system state, and the decoder outputs the surface roughness, vibration amplitude and geometric error corresponding to each prediction moment. Step 4: Imaginative Autonomous Decision Planning: Based on the preset action space sampling, multiple candidate processing action sequences are generated. In the forward dynamic world model, each candidate processing action sequence is deduced in multiple steps. The prediction results of each candidate sequence are evaluated and the best one is selected according to the preset multi-objective cost function. The optimal processing action sequence is output and its first step is taken as the current action to be issued. Step 5: Asynchronous Dual-Core Autonomous Execution with Dual Time Scales: The imagination and planning in Steps 3 and 4 are completed in the cognitive core running in the first cycle, and the constrained servo compensation is issued in the execution core running in the second cycle; the first cycle is longer than the second cycle, and the ratio of the first cycle to the second cycle satisfies the singular perturbation stability condition; the cognitive core and the execution core are asynchronous and parallel and share the system state; the execution core superimposes the currently issued action on the baseline process parameters and then issues it to the precision CNC machine tool in real time via the CNC system; Step Six: Embodied Loop Closure and Online Continuous Self-Learning: Collect the real-time measurement results of the machine tool, calculate the residual between it and the prediction results of the forward dynamic world model, use the residual to perform online continuous learning and fine-tuning of the forward dynamic world model and update the slowly varying hidden state; and reverse-anchor the surface roughness predicted in Step Three to the active Gaussian elements intersecting with the tool-workpiece contact area for pseudo-color rendering; return to Step Two to perform rolling loop closure.

[0014] Furthermore, in step one, the Gaussian element is represented by a 66-dimensional unified tensor, denoted as... Among them, the standard geometric and photometric properties comprise 59 dimensions, including three-dimensional position. Three-dimensional scale, four-dimensional rotational quaternion, one-dimensional opacity With the 48-dimensional spherical harmonic coefficient ; The physical sensing channel It is a seven-dimensional vector, and ,in: Scalar temperature; It is the root mean square vector of triaxial vibration. It is a three-axis cutting force vector; the physical sensing channel is divided into seven dimensions. To extend to higher dimensions as needed, the method is as follows: while keeping the mechanism of per-element anchoring and synchronous updating with geometric pose transformation unchanged, one or more spatially distributed physical field channels are added to each high-dimensional element. The spatially distributed physical field channels are selected from one or more of the ambient temperature field, the cutting fluid temperature field, and the spindle or guide rail oil temperature field.

[0015] Furthermore, in step one, the method for injecting computer-aided design CAD component-level semantic labels into the reconstructed Gaussian primitives is as follows: cross-modal registration is performed between the CAD triangular mesh model and the 3D Gaussian sputtering reconstructed point cloud from coarse to fine. The coarse registration is based on the Fast Point Feature Histogram (FPFH) and Random Sample Consistency (RANSAC) to solve the similarity transformation, while the fine registration uses the point-to-surface iterative nearest point (ICP). Then, the component labels of the CAD assembly tree are propagated to each Gaussian primitive using a KD tree combined with K-nearest neighbor search, and the Gaussian primitives are divided into load-bearing chains and non-load-bearing regions accordingly. In step one, the method for reconstructing the cutting force, vibration, and temperature signals into a continuous physical field using neural operators is as follows: Fourier neural operator (FNO) or deep operator network (DeepONet) is used to learn the operator mapping from the continuous physical field observed by sparse sensing to reconstruct the temperature field, vibration field, and cutting force field; and a graph neural network (GNN) is used to pass messages along the assembly topology on the load-bearing chain, so that the vibration and cutting force propagate along the kinematic chain and an impedance attenuation factor is introduced across the assembly interface to mask the non-load-bearing area. In step one, global or regional scalar parameters characterizing the processing environment state are also collected, and a global environment state vector is constructed. The global environment state vector This includes one or more of the following: relative humidity, air pressure, low-frequency environmental vibration, cutting fluid flow rate, cutting fluid pressure, and pneumatic system pressure; the global environmental state vector. Instead of anchoring each primitive, the hidden state encoding is incorporated in step two and used as a conditional input to the forward dynamic world model in step three.

[0016] In step one, the appearance mode, temperature mode, vibration mode and cutting force mode of the physically perceived three-dimensional Gaussian sputtering twin field share the same rasterization rendering pipeline. When switching modes, only the attribute buffer pointer is rebound without rebuilding the spatial acceleration structure.

[0017] Furthermore, in step two, the system state... It also includes a gradually changing hidden state characterizing tool wear or breakage, as well as the initial stress or surface hardness of the workpiece; the gradually changing hidden state is estimated online by cutting force signal characteristics, acoustic emission signals or visual inspection results, and continuously corrected by the residual in step six, serving as the conditional input of the forward dynamic world model; In step two, processing-related information is hierarchically carried according to its physical properties. The method is as follows: physical fields with spatial distribution characteristics are carried in the physically-sensing three-dimensional Gaussian sputtering twin field in a per-element anchoring manner; global or regional scalars without spatial distribution characteristics are carried in the global environment state vector. The tool and workpiece states, which are carried and used as conditions in the forward dynamic world model, vary gradually with the machining process and are carried and estimated online in the gradually changing hidden states.

[0018] Furthermore, in step three, the forward dynamic world model is composed of one or more of the following: the cyclic state-space model RSSM, the diffusion transformer operating in the latent space, or the joint embedded prediction architecture JEPA; the forward dynamic world model is trained with historical interaction data, which is composed of parametric simulation data and actual machine processing data, and the gap between simulation and reality is bridged by domain randomization. The training objective of the forward dynamic world model includes a physical consistency regularization term, which constrains the predicted state to satisfy physical priors such as cross-interface impedance attenuation, propagation along the kinematic chain, time-scale separation, and causal consistency between the environmental state and the processing state in the form of the residual of a neural operator or a physical information neural network PINN.

[0019] Furthermore, in step four, the candidate processing action sequence is sampled and extrapolated in parallel using the Model Predicted Path Integral (MPPI), Cross-Entropy Method (CEM), or diffusion strategy. The multi-objective cost function is a weighted sum of predicted surface roughness, predicted vibration amplitude, predicted geometric error, energy consumption and time penalty terms, and environmental robustness penalty terms. The environmental robustness penalty term increases the cost when the extrapolated environmental state or extended physical field in the next few steps exceeds a preset safety threshold. Step four is repeated in a rolling time domain manner to achieve rolling replanning. The forward dynamic world model is a causal world model that intervenes in the causal relationship and uses counterfactual imagination on candidate processing actions, and enables the imaginative autonomous decision-making planning to prioritize robust processing action sequences that are insensitive to environmental disturbances and operating condition drift.

[0020] Furthermore, in step five, the execution kernel performs finite impulse response (FIR) exponential decay low-pass filtering on the acquired raw signal, and uses the filtered root mean square amplitude of the vibration and its rate of change to form a state vector; and under the guidance of the action target given by the cognitive kernel, a multi-variable collaborative compensation quantity is generated by the compensation law subject to saturation and rate constraints. The multi-variable collaborative compensation quantity is the feed rate and spindle speed, or two or more of the grinding wheel linear speed, workpiece speed, axial feed and grinding depth, wherein the high-frequency channel is responsible for rapid suppression and the low-frequency channel is responsible for depth adjustment.

[0021] This invention also proposes a system for constructing and autonomously deciding on a world model of a precision CNC machine tool, used to realize the method for constructing and autonomously deciding on a world model of a precision CNC machine tool as described above, characterized by including: The multimodal embodied perception module is used to collect cutting force, vibration, temperature signals, multi-view images, and environmental perception signals characterizing the machining environment during the machine tool processing process. PA-3DGS twin construction module is used to reconstruct and build a physical perception 3D Gaussian sputtering twin field that is injected with CAD component-level semantics, reconstructed by neural operators, and anchored to the physical perception channel on a per-primary basis. The embedded hidden state encoding module is used to compress the twin field, multimodal signals and the environmental perception signals into a system state, and incorporate the tool and workpiece states that change slowly with the machining process as slowly changing hidden states into the system state. The forward dynamic world model module under action conditions is used to imagine and predict the future system state, surface roughness, vibration amplitude and geometric error based on candidate processing actions and the environmental perception signals. The imaginative autonomous decision-making and planning module is used to deduce and select the best processing action from the candidate processing action sequence and output the optimal processing action; The dual-timescale asynchronous dual-core execution module includes an asynchronous parallel cognitive core and an execution core that share the system state, used to complete imaginary planning and constrained servo compensation delivery under the condition of satisfying singular perturbation stability. The embodied closed-loop online continuous self-learning module is used to fine-tune the world model, update the slowly varying hidden states, and back-anchor the prediction results to the rendering by continuously learning online using prediction-measurement residuals.

[0022] Furthermore, a vision-language-action (VLA) task interface is provided, which uses natural language or task-level instructions to set the weights of the processing objective and the multi-objective cost function, and is then resolved into a planning objective by the embodied world model.

[0023] This invention also proposes an embodied intelligent system, including multiple precision CNC machine tool embodied world models as described above, and autonomous decision-making systems. Each of these precision CNC machine tool embodied world models serves as a machine tool-level single machine body. The machine tool-level single machine body undergoes multi-machine collaborative imagination and scheduling through a workshop-level embodied world model, and interacts bidirectionally with the basic model layer of the cloud-based manufacturing world model through federated continuous learning, knowledge distillation, and the model layer, in order to achieve collaborative forward-looking deduction and autonomous decision-making of the coupled state of workpiece, machine tool, and logistics.

[0024] The beneficial effects of this invention are as follows: The present invention provides a method for constructing and autonomously making decisions on the embodied world model of precision CNC machine tools, achieving a qualitative leap from mirror reproduction to forward-looking deduction and autonomous decision-making, and obtaining the following technical effects. (1) Forward-looking simulation and proactive decision-making. Steps three and four use the forward dynamic world model of action conditions and imaginative autonomous decision-making planning to enable the system to pre-simulate the future processing state (surface roughness, vibration amplitude, geometric error) of multiple candidate action sequences without moving the tool. The system then selects the best action according to the multi-objective cost function, realizing virtual trial and error and physical execution. This fundamentally overcomes the shortcomings of existing digital twins, which can only reproduce history and existing controls, which can only react passively.

[0025] (2) Balancing physical reliability and real-time performance. Step 1 uses Physically Aware 3D Gaussian Sputtered Twin Field (PA-3DGS) to reconstruct physical fields such as force, vibration, and temperature through neural operators and anchors them to Gaussian primitives one by one. The twin model is updated synchronously with geometric pose changes, making the twin model "what you see is what you get, and what you get is what you feel." Combined with physical consistency regularization constraints, it avoids the blind physics of purely data-driven models. At the same time, the four modes of appearance, temperature, vibration, and force share the same rasterization rendering pipeline. When switching modes, only the attribute buffer pointer is rebound, reducing latency and maintaining high frame rate real-time rendering.

[0026] (3) Resolving time mismatch. The dual-timescale asynchronous dual-core architecture in step five decouples the cognitive core and the execution core. The timescale separation ratio satisfies the singular perturbation stability condition, eliminating the dynamic interference between deep prediction latency and high-frequency servo execution.

[0027] (4) Closed-loop self-evolution step six: Driven by prediction-measurement residuals, online continuous learning is used to fine-tune the world model and update the slowly changing hidden state, forming a self-consistent closed loop of "perception-imagination-planning-execution-self-learning", which enables the model to have the ability to self-correct and continuously evolve.

[0028] (5) Information is carried in a hierarchical manner. The spatially distributed physical field (anchored to PA-3DGS), the global environmental scalar (as the condition input of the world model), and the tool-workpiece gradually changing hidden state (online estimation) are carried in three layers respectively, so that the world model is more complete and not redundant.

[0029] This invention constructs a physically perceptible three-dimensional Gaussian sputtering twin field as a unified representation base, combines a forward dynamic world model of action conditions with imaginative autonomous decision-making and planning, and employs a dual-timescale asynchronous dual-core architecture to achieve forward-looking deduction, autonomous decision-making, and closed-loop self-learning of the machining process. This fundamentally solves the problems of lack of forward-looking prediction, passive control response, and time mismatch in existing technologies, making precision machine tools into embodied intelligent agents capable of forward-looking deduction and autonomous decision-making. It can be extended to various precision manufacturing scenarios such as milling, grinding, and gear machining. Attached Figure Description

[0030] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the following figures are provided for illustration: Figure 1 This is a flowchart illustrating the construction and autonomous decision-making method of the embodied world model of a precision CNC machine tool according to the present invention. Figure 2 Schematic diagram of the construction of scalable physical perception 3D Gaussian tensor primitives; Figure 3 A structural diagram of the embodied world model of forward dynamics under action conditions; Figure 4 A flowchart for imaginative autonomous decision-making planning (MPPI / CEM / diffusion strategy); Figure 5 Flowchart of dual-time-scale asynchronous dual-core autonomous execution closed-loop control; Figure 6 Flowchart for online continuous self-learning (training during testing) and prediction back-anchoring rendering; Figure 7 This is the overall architecture diagram of the precision CNC machine tool embodied world model construction and autonomous decision-making system of the present invention; Figure 8 This is a schematic diagram of a layered embodied architecture for a future workshop-level embodied intelligence system. Detailed Implementation

[0031] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0032] This embodiment aims to propose a method, system, and embodied intelligent system for constructing an embodied world model and making autonomous decisions for precision CNC machine tools. It constructs a three-dimensional Gaussian sputtering twin field for physical perception, injecting semantics at the CAD component level, reconstructing it using neural operators, and anchoring physical perception channels element by element. This serves as a unified spatial anchoring representation base for the embodied world model. On this base, a forward dynamic world model conditioned by actions and an imaginative autonomous decision-making mechanism are established. A self-consistent closed loop of perception-imagination-planning-execution-self-learning is achieved using a dual-timescale asynchronous dual-core architecture under singular perturbation stability conditions. This allows the precision machine tool to leap from "mirror reproduction" to "forward-looking deduction and autonomous decision-making," solving technical problems such as the lack of foresight in existing digital twins, the inability of existing control systems to only react, and time mismatch between deep prediction and servo execution.

[0033] Specifically, such as Figure 1 As shown in the figure, the method for constructing and autonomously making decisions on the embodied world model of a precision CNC machine tool in this embodiment includes the following steps.

[0034] Step 1: Construction of Physics-Aware Embodied Representation: Multimodal sensors are deployed on the precision CNC machine tool and workpiece to collect cutting force, vibration, and temperature signals, as well as multi-view images during the machining process. Based on the multi-view images, the spatial geometry of the machine tool and workpiece is reconstructed using 3D Gaussian sputtering with an explicit radiation field. Computer-aided design (CAD) component-level semantic tags are injected into the reconstructed Gaussian primitives. The cutting force, vibration, and temperature signals are reconstructed into a continuous physical field using neural operators. The physical field values ​​at the center of each Gaussian primitive are then anchored to the corresponding Gaussian primitives, and the physical field values ​​are updated synchronously with the geometric pose changes of the Gaussian primitives. This results in a Physics-Aware 3D Gaussian sputtered twin field, denoted as PA-3DGS twin field, which augments the physical perception channels on the basis of standard geometry and photometric properties. This serves as the unified spatial anchoring representation base for the embodied world model.

[0035] like Figure 2 As shown, the spatial geometry of the machine tool and workpiece is reconstructed using three-dimensional Gaussian sputtering with explicit radiation fields from multi-view images. Each Gaussian element is characterized by an anisotropic three-dimensional Gaussian function, and its covariance matrix is ​​parameterized by the scale vector and rotation matrix to ensure positive semidefiniteness.

[0036] in: Let Gausky's covariance matrix be denoted by Gausky. Let the rotation matrix represent the rotational attitude of the Gaussian metaspace; Let be the scale diagonal matrix representing the anisotropic scale of the Gaussian element.

[0037] To suppress excessive cloning of primitives caused by reflective metallic surfaces, an adaptive density control strategy is adopted, which increases the cloning gradient threshold and tightens the splitting scale boundary. End-to-end optimization is performed using a weighted combination of L1 loss and structural similarity D-SSIM loss. In this embodiment, Gaussian primitives are represented using a 66-dimensional unified tensor, denoted as […]. Among them, the standard geometric and photometric properties comprise 59 dimensions, including three-dimensional position. Three-dimensional scale, four-dimensional rotational quaternion, one-dimensional opacity With the 48-dimensional spherical harmonic coefficient Specifically, physical sensing channels It is a seven-dimensional vector, and ,in: Scalar temperature; It is the root mean square vector of triaxial vibration. This is the three-axis cutting force vector.

[0038] In this embodiment, the physical sensing channel is also divided into seven dimensions. To extend to higher dimensions as needed, the method is as follows: while keeping the mechanism of per-element anchoring and synchronous updating with geometric pose transformation unchanged, one or more spatially distributed physical field channels are added to each high-dimensional element. The spatially distributed physical field channels are selected from one or more of the ambient temperature field, the cutting fluid temperature field, and the spindle or guide rail oil temperature field.

[0039] To imbue the twin field with component-level engineering semantics, a coarse-to-fine cross-modal registration is performed between the CAD triangular mesh model and the reconstructed point cloud. In this embodiment, the method for injecting computer-aided design CAD component-level semantic labels into the reconstructed Gaussian primitives is as follows: A coarse-to-fine cross-modal registration is performed between the CAD triangular mesh model and the 3D Gaussian sputtering reconstructed point cloud. The coarse registration is based on Fast Feature Histogram (FPFH) and Random Sample Consistency (RANSAC) to solve the similarity transformation, while the fine registration uses iterative nearest point (ICP) from point to surface. Then, using a KD-tree combined with K-nearest neighbor search, the component labels from the CAD assembly tree are propagated to each Gaussian primitive, and the structural boundaries of the load-bearing chain and non-load-bearing region are established accordingly. The Gaussian primitives are divided into the load-bearing chain and the non-load-bearing region, serving as the basis for subsequent differentiated physical field mapping. Specifically, the load-bearing chain is the set of Gaussian primitives along the mechanical structure topology path that continuously transmits cutting force and vibration from the tool-workpiece contact area through the workpiece, fixture, worktable, and spindle. The non-load-bearing region consists of the remaining Gaussian primitives outside the load-bearing chain.

[0040] To address the ill-conditioned inverse problem caused by sparse sensing points, this embodiment does not employ static interpolation or offline finite element method, but instead uses neural operators to reconstruct the continuous physical field. In this embodiment, the method for reconstructing the cutting force, vibration, and temperature signals into a continuous physical field using neural operators is as follows: Fourier neural operator (FNO) or deep operator network (DeepONet) is used to learn the operator mapping from sparse sensing observations (such as triaxial force measurement, triaxial acceleration, and multi-point temperature measurement) to the continuous physical field, obtaining the physical field function that evolves over time. The temperature field, vibration field, and cutting force field are reconstructed, and a graph neural network (GNN) is used along the assembly topology on the load-bearing chain to pass messages, allowing vibration and cutting force to propagate along the kinematic chain and introducing an impedance attenuation factor across the assembly interface, thus masking the non-load-bearing area. Specifically, the physical field function that evolves over time is expressed as:

[0041] in: It is a combination tensor of continuous physical field functions that evolves over time; This represents the continuous scalar temperature field function value at that spatial location and time. The root mean square vector function value of continuous triaxial vibration at this spatial location and time; This represents the continuous triaxial cutting force vector function value at this spatial position and time. , , and These represent the three spatial coordinate components and the time variable of a three-dimensional Cartesian coordinate system.

[0042] The temperature field is reconstructed using operators with physical boundary constraints, with key heat sources and ambient cold sources serving as boundary anchors. The vibration and cutting force fields are propagated along the assembly topology using a graph neural network (GNN) on the load-bearing chain. The propagation path exhibits exponential decay, and an impedance attenuation factor is introduced across the assembly interface to mask the cutting force channels in non-load-bearing areas. The neural operator reconstruction is differentiable, facilitating joint training with the world model and embedding into real-time loops.

[0043] In this embodiment, anchoring is defined as: for each center located at The Gaussian primitive is assigned the field value at its center to the physical sensing channel of that primitive, i.e. When the machine tool feed axis moves or the workpiece rotates with the rotary axis, the basic element undergoes a rigid body pose transformation. , When (R is the rotation matrix, d is the translation vector), the physical sensing channel is updated synchronously: the primitive center is pressed... Update; the scalar temperature channel is rotationally invariant, and is only resampled at the updated position, i.e. The triaxial vibration channel and the triaxial cutting force channel are vectors, and they need to be rotated in the same way to maintain a geometrically consistent reference frame. , For non-load-bearing primitives, their cutting force channels are zeroed according to the zero-masking rule. Thus, after the sparse sensing signal is reconstructed by the neural operator continuous field, it is precisely and in real time anchored to hundreds of thousands of Gaussian primitives in the manner of "sampling at the center of each primitive and performing rigid body transformation with pose", and maintains the consistency of physical field and spatial geometry with toolpath movement and workpiece rotation.

[0044] This embodiment also collects global or regional scalar parameters characterizing the processing environment state and constructs a global environment state vector. The global environment state vector This includes one or more of the following: relative humidity, air pressure, low-frequency environmental vibration, cutting fluid flow rate, cutting fluid pressure, and pneumatic system pressure; the global environmental state vector. Instead of anchoring each primitive, the hidden state encoding is incorporated in step two and used as a conditional input to the forward dynamic world model in step three.

[0045] In this embodiment, the appearance mode, temperature mode, vibration mode and cutting force mode of the physically perceived three-dimensional Gaussian sputtering twin field share the same rasterization rendering pipeline. When switching modes, only the attribute buffer pointer is rebound without rebuilding the spatial acceleration structure, so that the four-mode switching latency is less than 16 milliseconds.

[0046] Specifically, in this embodiment, the physical sensing channel can be expanded as needed, and processing-related information is carried hierarchically according to its physical attributes (see...). Figure 8 (As shown in the lower half): Firstly, for extended physical quantities with spatial distribution characteristics (such as ambient temperature field, cutting fluid temperature field, spindle or guide rail oil temperature field), as newly added spatially distributed physical field channels, the same "neural operator reconstruction - sampling per elementary center - updating with pose rigid body transformation" mechanism as mentioned above is used for anchoring, so that the physical perception channel is transformed from... First, it extends to higher dimensions while maintaining unified rasterization and lightweight characteristics; second, for global or regional scalars without spatial distribution characteristics, it does not perform element-by-element anchoring, but constructs a global environment state vector. (These can be relative humidity, air pressure, low-frequency environmental vibration, cutting fluid flow rate, cutting fluid pressure, and pneumatic system pressure, etc.), serving only as conditions for the world model; thirdly, for the tool and workpiece states that change gradually with machining, a gradually changing hidden state is constructed and incorporated into the system state, estimated online by cutting force characteristics, acoustic emission, or visual inspection, and continuously corrected by the residuals in step six. Each of the three types of information is properly positioned, both expanding the completeness of the world model and avoiding redundancy and inconsistency.

[0047] Step 2: Embodied Latent State Encoding: The physically-aware 3D Gaussian sputtering twin field (PA-3DGS twin field) and multimodal signals are compressed into a compact implicit representation using an encoder (e.g., a 3D variational autoencoder or a joint embedded encoder), and then combined with the tool pose, current process parameters, and a global environment state vector characterizing the machining environment state. And by splicing together the gradually changing hidden states that characterize the gradually changing properties of the tool and the workpiece, the system state at the current moment is obtained. Hidden states, while compressing redundancy, retain cross-modal coupling information and environmental / condition context, providing low-dimensional input for efficient imagination of the world model.

[0048] In this embodiment, the system state It also includes a gradually changing hidden state characterizing tool wear or breakage, as well as the initial stress or surface hardness of the workpiece; the gradually changing hidden state is estimated online by cutting force signal characteristics, acoustic emission signals or visual inspection results, and continuously corrected by the residual in step six, serving as the conditional input of the forward dynamic world model.

[0049] In this embodiment, processing-related information is hierarchically carried according to its physical properties. The method is as follows: physical fields with spatial distribution characteristics are carried in the physically-aware three-dimensional Gaussian sputtering twin field (PA-3DGS twin field) in a per-element anchoring manner; global or regional scalars without spatial distribution characteristics are carried in the global environment state vector. The tool and workpiece states, which are carried and used as conditions in the forward dynamic world model, vary gradually with the machining process and are carried and estimated online in the gradually changing hidden states.

[0050] Step 3: Forward Dynamics Imagining of Action Conditions: Using candidate machining actions, the global environment state vector, and the slowly changing hidden state as conditions, the learned forward dynamics world model is used to recursively imagine and predict the system state at the next moment and even in the future H steps from the current system state. The decoder outputs the surface roughness, vibration amplitude, and geometric error corresponding to each prediction moment. In this embodiment, candidate machining actions include at least the feed rate increment, spindle or workpiece rotation speed increment, and tool path increment.

[0051] like Figure 3 As shown, the forward dynamics world model uses candidate processing actions (Including feed rate increments, spindle or workpiece speed increments, and toolpath increments) and global environment vectors Conditioned by the slowly changing hidden state, the current system state Recursive imagination prediction:

[0052] in: Let t be the slowly varying hidden state vector representing tool wear and workpiece state at time t; To parameterize The forward dynamic state transition function (or transition network).

[0053] By recursively deriving H steps forward, we obtain ,..., The decoder outputs the predicted surface roughness, predicted vibration amplitude, and predicted geometric error corresponding to the hidden state at each time step.

[0054] in: , and These are the predicted surface roughness, vibration amplitude, and geometric error, respectively. To parameterize The feature decoding header function (or decoding network).

[0055] One implementation of the forward dynamical world model is a combination of a 3D variational autoencoder and a latent space diffusion transformer: the autoencoder encodes Gaussian units into a compact latent space, and the diffusion transformer iteratively "imagines" future latent states on the latent patch conditioned by actions, environmental conditions, and denoising steps; another implementation is the cyclic state-space model RSSM; yet another implementation is the joint embedding prediction architecture JEPA, which performs predictions in the representation space rather than the pixel space to improve robustness and efficiency.

[0056] In this embodiment, the forward dynamics world model is pre-trained offline and fine-tuned online using historical interaction data. This historical interaction data comprises parametric simulation data and data collected from actual machine variable-parameter machining. Domain randomization is used to bridge the gap between simulation and reality. In this embodiment, the parametric simulation data is generated from parametric cutting / grinding simulations, covering force, vibration, temperature, and surface morphology under different toolpaths, process parameters, and environmental conditions. To bridge the gap between simulation and reality, a hybrid training strategy of domain randomization and "simulation pre-training followed by actual machine fine-tuning" is adopted, with the weight of actual machine data gradually increased as the machining process progresses.

[0057] To ensure physical reliability, the training objective of the forward dynamics world model includes a physical consistency regularization term. This term, in the form of a neural operator or a physical information neural network PINN, constrains the predicted state to satisfy physical priors such as cross-interface impedance attenuation, propagation along the kinematic chain, time-scale separation, and causal consistency between the environmental state and the processing state. This overcomes the physical blindness of purely data-driven models.

[0058] Step 4: Imaginative Autonomous Decision Planning: Based on the preset action space sampling, multiple candidate action sequences are generated. In the forward dynamic world model, each candidate action sequence is deduced in multiple steps. The prediction results of each candidate sequence are evaluated and the best one is selected according to the preset multi-objective cost function. The optimal action sequence is output and its first step is taken as the current action to be issued.

[0059] like Figure 4 As shown, in the learned forward dynamic world model, K candidate processing action sequences are sampled and parallel multi-step extrapolation (virtual trial and error) can be performed. Model prediction path integral (MPPI), cross-entropy method (CEM), or diffusion strategy can be used. The multi-objective cost function is a weighted sum of predicted surface roughness, predicted vibration amplitude, predicted geometric error, energy consumption and time penalty terms, and environmental robustness penalty terms. The environmental robustness penalty term increases the cost when the environmental state or extended physical field in the extrapolated future steps exceeds a preset safety threshold. Step four is repeated in a rolling time domain manner to achieve rolling replanning.

[0060] Specifically, for each candidate sequence, the future prediction given by the decoder head is calculated using a multi-objective cost function. Conduct an assessment:

[0061] in: , , These are the predicted surface roughness, vibration amplitude, and geometric error, respectively. , Penalties for energy consumption and working hours; As an environmental robustness penalty, it increases the cost when the environmental state or extended physical field in the future exceeds the preset safety threshold, so that the planning will prioritize action sequences that are not sensitive to environmental disturbances. to These are weighting coefficients. Based on MPPI... The optimal processing action sequence is obtained by exponential weighted fusion or CEM elite resampling, and the first step is taken as the current action to be issued; this process is repeated in a rolling time domain manner to achieve rolling replanning.

[0062] Furthermore, the forward dynamics world model can serve as a causal world model to intervene causally and imagine counterfactual scenarios for candidate actions, enabling the imagined autonomous decision-making planning to prioritize robust machining action sequences that are insensitive to environmental disturbances and operating condition drift, thereby selecting the most robust action under various potential operating conditions. Unlike existing controls that only respond to the current error, this step pre-rehearses and autonomously selects the optimal action before the error is irreversibly etched into the workpiece.

[0063] Step 5: Asynchronous Dual-Core Autonomous Execution with Dual Time Scales: The imagination and planning in Steps 3 and 4 are completed in the cognitive core running in the first cycle, and the constrained servo compensation is issued in the execution core running in the second cycle; the first cycle is longer than the second cycle, and the ratio of the first cycle to the second cycle satisfies the singular perturbation stability condition; the cognitive core and the execution core are asynchronous and parallel and share the system state; the execution core superimposes the currently issued action on the baseline process parameters and then issues it to the precision CNC machine tool in real time via the CNC system.

[0064] In this embodiment, the acquired raw signal is subjected to finite impulse response (FIR) exponential decay low-pass filtering, and the filtered root mean square amplitude of the vibration and its rate of change constitute a state vector. Guided by the action target given by the cognitive kernel, a multivariate collaborative compensation quantity is generated under the saturation and rate constraints of the compensation law. The multivariate collaborative compensation quantity is the feed rate and spindle speed, or two or more of the following: grinding wheel linear speed, workpiece speed, axial feed, and grinding depth. The high-frequency channel is responsible for rapid suppression, and the low-frequency channel is responsible for depth adjustment. Specifically, in this embodiment, the first period is 300 to 400 milliseconds, the second period is 10 to 20 milliseconds, and the ratio of the first period to the second period is 15 to 30 to satisfy the singular perturbation stability condition.

[0065] like Figure 5 As shown, computationally intensive steps three and four are placed in the first cycle. The running cognitive core places lightweight servo compensation in the second cycle. The running execution cores are asynchronous and run in parallel, sharing the system state in the same PA-3DGS space anchoring buffer. In one embodiment... Approximately 350 milliseconds The time interval is approximately 15 milliseconds, and the time-scale separation ratio is approximately 23, satisfying the singular perturbation stability condition, thereby eliminating dynamic interference between depth prediction delay and high-frequency servo execution. The execution kernel performs an FIR exponentially decaying low-pass filter on the original signal, and the filtered root mean square amplitude of the vibration and its rate of change constitute the state vector. Guided by the action target given by the cognitive kernel, the compensation law, subject to saturation and rate constraints, generates multi-variable collaborative compensation, which is then superimposed on the baseline parameters and sent to the CNC system. For example: , ;in: This is the spindle speed command value sent to the CNC system after compensation; Baseline process parameters (or initial settings) for spindle speed; The gain coefficient for coordinating the spindle speed; The target adjustment amount of the spindle speed is generated under the guidance of the cognitive kernel; This is the feed rate command value sent to the CNC system after compensation; The baseline process parameter (or initial setting) for the feed rate; The gain coefficient for the coordinated compensation of feed rate; The target adjustment amount of feed rate is generated under the guidance of the cognitive kernel. For milling, the multivariate is the feed rate and the spindle speed, with the feed rate responsible for high-frequency rapid suppression and the spindle speed responsible for low-frequency depth adjustment; for grinding / gear machining, the multivariate can be replaced by two or more of the following: grinding wheel linear speed, workpiece speed, axial feed, and grinding depth.

[0066] Furthermore, the above design satisfies the time-scale separation condition for singular perturbations in an engineering sense: the machine tool servo drive, as a fast variable, has a time constant of approximately 1 to 5 milliseconds; the chatter setup, as a slow variable, has a characteristic time of approximately 100 to 500 milliseconds. By setting the execution kernel cycle to the same order of magnitude as the servo time constant, and the cognitive kernel cycle to be much larger than the servo time constant and comparable to the chatter characteristic time, and taking a time-scale separation ratio of 15 to 30, the fast subsystem has fully converged within one update cycle of the slow subsystem, and the slow subsystem can be considered as a quasi-steady state. This invention further ensures the bounded response of the fast subsystem by applying saturation and rate constraints to the compensation law and introducing a physical consistency regularization term for time-scale separation in the world model training objective, preventing divergence or oscillation due to time delay mismatch when the cognitive and execution kernels are asynchronously parallel.

[0067] Step Six: Embodied Loop Closure and Online Continuous Self-Learning: Collect real-time measurement results from the machine tool, calculate the residual between the residual and the prediction results of the forward dynamic world model, use the residual to perform online continuous learning fine-tuning of the forward dynamic world model and update the slowly varying hidden state; and back-anchor the surface roughness predicted in Step Three to the active Gaussian elements intersecting the tool-workpiece contact area for pseudo-color rendering; return to Step Two to perform rolling loop closure. In this embodiment, online continuous learning fine-tuning adopts test-time training and uses a continuous learning strategy including elastic weight consolidation to constrain weight updates to suppress catastrophic forgetting. Fine-tuning is triggered once after each part is machined or at preset intervals.

[0068] like Figure 6 As shown, real-time measurement results of the machine tool are collected (such as in-machine inspection or post-measurement feedback), and the residual Δ between the results and the world model prediction is calculated. This residual is then used for online fine-tuning of the forward dynamic world model during testing. Simultaneously, to suppress catastrophic forgetting, continuous learning strategies such as elastic weight consolidation are employed to constrain weight updates. The training objective can be expressed as:

[0069] in: Total training loss. This is for predicting the residual loss term. For physical consistency regularization; For the Elastic Weight Consolidation (EWC) penalty term, For the first Diagonal elements of the Fisher information matrix with parameters. These are the current training parameters. This parameter is optimized for historical data. It constrains core weight updates and suppresses catastrophic forgetting. and The weighting coefficients for the regularization and anti-forgetting terms are used to balance the contributions of each loss term.

[0070] Simultaneously, the gradually changing hidden states (such as tool wear estimation) are updated based on the residual trend, allowing the prediction to be continuously calibrated as the machining process progresses, forming a correctable and sustainably evolving embodied world model. The surface roughness predicted in step three is then back-anchored to the active Gaussian elements intersecting the tool-workpiece contact area under the current tool pose, serving as an additional transient attribute and rendered with pseudo-color to represent the degradation trend. This achieves a unified synchronization of "prediction-space-vision" without increasing geometric registration overhead. The four modes—appearance, temperature, vibration, and force—share the same rasterization pipeline. During mode switching, only the attribute buffer pointers are rebound without rebuilding the spatial acceleration structure, resulting in a four-mode switching latency of less than 16 milliseconds. Finally, the process returns to step two for a rolling closed loop, forming a self-consistent closed loop of "perception-imagination-planning-execution-self-learning."

[0071] In one embodiment of this example, the device-level twin contains approximately 660,000 Gaussian elements, with a new perspective synthesis peak signal-to-noise ratio of approximately 30.94 dB, a structural similarity of approximately 0.9178, and a steady-state rendering frame rate of no less than 120 FPS. Under the same nominal toolpath, comparing the uncompensated baseline with the in-loop autonomous machining of this embodiment: the baseline exhibits non-uniform surface roughness along the machining direction and is accompanied by visible chatter textures. After adopting the method of this embodiment, the surface roughness is compressed into a uniform range, the degradation error is constrained to ΔRa≤0.13 μm, the texture is uniform and there are no chatter stripes, and the machining efficiency is not lost under the same number of measurement points.

[0072] Furthermore, this embodiment can set a vision-language-action (VLA) task interface: the operator sets the weights of the machining target and the multi-objective cost function using natural language or task-level instructions (e.g., "prioritize efficiency while ensuring Ra≤0.2μm"). The embodied world model then parses the result into the planning target for step four, enabling the precision machine tool to have task-level autonomous decision-making capabilities.

[0073] This embodiment also proposes a system for constructing and autonomously deciding on a world model of a precision CNC machine tool, used to implement the method described above for constructing and autonomously deciding on a world model of a precision CNC machine tool. Specifically, as follows... Figure 7 As shown, the precision CNC machine tool embodied world model construction and autonomous decision-making system in this embodiment includes a multimodal embodied perception module, a PA-3DGS twin construction module, an embodied hidden state encoding module, a motion condition forward dynamics world model module, an imaginative autonomous decision-making and planning module, a dual-timescale asynchronous dual-core execution module, and an embodied closed-loop online continuous self-learning module.

[0074] Specifically, the multimodal embodied perception module is used to collect cutting force, vibration, temperature signals, multi-view images, and environmental perception signals characterizing the machining environment during the machine tool machining process; the PA-3DGS twin construction module is used to reconstruct and construct a physical perception 3D Gaussian sputtering twin field that is injected with CAD component-level semantics, reconstructed by neural operators, and anchored to the physical perception channel element by element; the embodied hidden state encoding module is used to compress the twin field, multimodal signals, and the environmental perception signals into a system state, and incorporate the gradually changing tool and workpiece states as gradually changing hidden states into the system state; the action condition forward dynamics world model module is used to... The system uses candidate processing actions and environmental perception signals as conditions to imagine and predict future system states, surface roughness, vibration amplitude, and geometric errors. An imaginative autonomous decision-making and planning module is used to deduce and select the optimal processing action sequence from the candidate processing actions and output the optimal processing action. A dual-timescale asynchronous dual-core execution module includes an asynchronous parallel cognitive core and an execution core that share the system state. This module is used to complete imaginative planning and constrained servo compensation under singular perturbation stability conditions. An embodied closed-loop online continuous self-learning module is used to fine-tune the world model, update the slowly varying hidden state, and back-anchor the prediction results to the rendering using online continuous learning of prediction-measurement residuals. In this embodiment, the cognitive core and execution core are asynchronous parallel computing threads. They interact with the system state through a shared Gaussian meta-buffer in memory and interact with the CNC system and multimodal sensors through a bidirectional bus.

[0075] Furthermore, the precision CNC machine tool embodied world model construction and autonomous decision-making system in this embodiment is also equipped with a vision-language-action (VLA) task interface, which sets the weights of the machining target and the multi-objective cost function using natural language or task-level instructions, and is then parsed into planning targets by the embodied world model.

[0076] This embodiment also proposes an embodied intelligence system, including multiple precision CNC machine tool embodied world model construction and autonomous decision-making systems as described above. Each precision CNC machine tool embodied world model construction and autonomous decision-making system serves as a machine tool-level single machine body. The machine tool-level single machine body is subjected to multi-machine collaborative imagination and scheduling through the workshop-level embodied world model, and bidirectional interaction is carried out through federated continuous learning and knowledge distillation with the basic model layer of the cloud manufacturing world model to achieve collaborative forward-looking inference and autonomous decision-making of the coupled state of workpiece, machine tool and logistics.

[0077] Specifically, such as Figure 8 As shown, this embodiment can be expanded from a single machine to a layered embodied intelligence system for machining workshops and even across factories: each precision machine tool and the system of this invention running on it constitute a machine tool-level single machine tool body; multiple single machine tool bodies are collaboratively imagined and scheduled through a workshop-level embodied world model, and collaboratively predictive extrapolate the workpiece flow, machine tool occupancy, and logistics coupling status; each single machine tool body interacts bidirectionally with the workshop-level world model and the cloud-based manufacturing world model base model layer through federated continuous learning and knowledge distillation, aggregating group experience and distributing distilled knowledge while protecting the data of all parties, making the entire machining workshop a self-aware, self-imagining, autonomously decision-making, and continuously evolving embodied intelligent system. This layered embodied architecture is a future-oriented extended embodiment of this invention, and its single-machine layer can already be implemented independently.

[0078] The method, system, and embodied intelligent system for constructing an embodied world model of a precision CNC machine tool and making autonomous decisions have the following technical effects.

[0079] (1) From “mirror reproduction” to “forward-looking simulation and autonomous decision-making”. The forward dynamic world model based on action can predict the future processing state and surface quality without actually cutting, and based on this, it can make imaginative planning and autonomous selection, realize “virtual trial and error, physical execution”, and overcome the fundamental limitations of existing digital twins that can only reflect the present and the past, and existing control that can only react passively.

[0080] (2) Balancing physical reliability and real-time performance. By reconstructing the physical perception 3D Gaussian sputtering twin field and physical consistency regularization term through neural operators, physical boundaries and transmission laws are injected into the world model, avoiding the physical blindness of pure data models; by decoupling the shared rasterization pipeline for rendering, the multimodal switching latency and primitive scale are decoupled, maintaining high frame rate real-time visualization.

[0081] (3) Resolve time mismatch. The dual-timescale asynchronous dual-core architecture decouples the cognitive imagination-planning at the level of hundreds of milliseconds from the servo execution at the level of milliseconds. The timescale separation ratio satisfies the singular perturbation stability condition, fundamentally eliminating the dynamic interference between deep prediction latency and high-frequency servo execution.

[0082] (4) Self-correcting and continuous evolution. The model is trained during testing by predicting-measuring residuals and suppresses forgetting through continuous learning, forming a self-correcting and continuously evolving embodied world model; the prediction results are back-anchored for rendering, achieving a three-in-one synchronization of "prediction-space-vision".

[0083] (5) Information is hierarchically carried and scalable to the workshop level. Processing-related information is carried in three layers: spatially distributed physical field, global environmental scalar, and slowly varying hidden state, making the world model more complete and less redundant. It can be extended to workshop-level and even cross-plant-level embodied intelligence through hierarchical embodied architecture and federated continuous learning. Ultimately, without increasing redundant passes and offline measurements, the surface quality degradation of precision machining is suppressed to a very small range (e.g., ΔRa≤0.13μm), without sacrificing servo responsiveness and processing efficiency. It has good versatility and scalability and can be extended to various precision manufacturing scenarios such as milling, grinding, and gear processing.

[0084] The above-described embodiments are merely preferred embodiments provided to fully illustrate the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are all within the scope of protection of the present invention. The scope of protection of the present invention is defined by the claims.

Claims

1. A method for constructing an embodied world model and making autonomous decisions for a precision CNC machine tool, characterized in that: Includes the following steps: Step 1: Construction of Physically Perceived Embodied Representation: Multimodal sensors are deployed on a precision CNC machine tool and workpiece to collect cutting force, vibration, and temperature signals, as well as multi-view images during the machining process. Based on the multi-view images, the spatial geometry of the machine tool and workpiece is reconstructed using 3D Gaussian sputtering with an explicit radiation field. Computer-aided design (CAD) component-level semantic tags are then injected into the reconstructed Gaussian primitives. The cutting force, vibration, and temperature signals are reconstructed into a continuous physical field using neural operators. The physical field values ​​at the center of each Gaussian primitive are then anchored to the corresponding primitive, and these values ​​are updated synchronously with the geometric pose changes of the Gaussian primitives. This results in a physically perceived 3D Gaussian sputtered twin field with augmented physical perception channels based on standard geometry and photometric properties, serving as the unified spatial anchoring representation base for the embodied world model. Step 2: Embodied Hidden State Encoding: The physically perceived 3D Gaussian sputtering twin field and multimodal signal are compressed into a compact hidden representation by an encoder, and then concatenated with the tool pose, current process parameters, global environment state vector representing the machining environment state, and slowly varying hidden state representing the slowly varying characteristics of the tool and workpiece to obtain the system state at the current moment. Step 3: Forward dynamics imagination of action conditions: Using the candidate processing action, the global environment state vector and the slowly changing hidden state as conditions, the learned forward dynamics world model is used to recursively imagine and predict the system state at the next moment and even the next H steps from the current system state, and the decoder outputs the surface roughness, vibration amplitude and geometric error corresponding to each prediction moment. Step 4: Imaginative Autonomous Decision Planning: Based on the preset action space sampling, multiple candidate processing action sequences are generated. In the forward dynamic world model, each candidate processing action sequence is deduced in multiple steps. The prediction results of each candidate sequence are evaluated and the best one is selected according to the preset multi-objective cost function. The optimal processing action sequence is output and its first step is taken as the current action to be issued. Step 5: Asynchronous dual-core autonomous execution with dual time scales: The imagination and planning in Steps 3 and 4 are completed in the cognitive core running in the first cycle, and the constrained servo compensation is distributed to the execution core running in the second cycle; the first cycle is longer than the second cycle, and the ratio of the first cycle to the second cycle satisfies the singular perturbation stability condition; The cognitive kernel and the execution kernel operate asynchronously in parallel and share the system state; The execution core superimposes the current action onto the baseline process parameters and then sends it to the precision CNC machine tool in real time via the CNC system. Step Six: Embodied Loop and Online Continuous Self-Learning: Collect the real-time measurement results of the machine tool, calculate the residual between the residual and the prediction results of the forward dynamic world model, and use the residual to perform online continuous learning and fine-tuning of the forward dynamic world model and update the slowly changing hidden state; The surface roughness predicted in step three is then reverse-anchored to the active Gaussian elements intersecting the tool-workpiece contact area for pseudo-color rendering; the process then returns to step two for rolling loop closure.

2. The method for constructing and autonomously deciding on the embodied world model of a precision CNC machine tool according to claim 1, characterized in that: In step one, the Gaussian element is represented by a 66-dimensional unified tensor, denoted as Among them, the standard geometric and photometric properties comprise 59 dimensions, including three-dimensional position. Three-dimensional scale, four-dimensional rotational quaternion, one-dimensional opacity With the 48-dimensional spherical harmonic coefficient ; The physical sensing channel It is a seven-dimensional vector, and ,in: Scalar temperature; It is the root mean square vector of triaxial vibration. It is a three-axis cutting force vector; the physical sensing channel is divided into seven dimensions. To extend to higher dimensions as needed, the method is as follows: while keeping the mechanism of per-element anchoring and synchronous updating with geometric pose transformation unchanged, one or more spatially distributed physical field channels are added to each high-dimensional element. The spatially distributed physical field channels are selected from one or more of the ambient temperature field, the cutting fluid temperature field, and the spindle or guide rail oil temperature field.

3. The method for constructing and autonomously deciding on the embodied world model of a precision CNC machine tool according to claim 1, characterized in that: In step one, the method for injecting computer-aided design CAD component-level semantic labels into the reconstructed Gaussian primitives is as follows: cross-modal registration is performed between the CAD triangular mesh model and the 3D Gaussian sputtering reconstructed point cloud from coarse to fine. The coarse registration is based on the Fast Point Feature Histogram (FPFH) and Random Sample Consistency (RANSAC) to solve the similarity transformation, while the fine registration uses the point-to-surface iterative nearest point (ICP). Then, the component labels of the CAD assembly tree are propagated to each Gaussian primitive using a KD tree combined with K-nearest neighbor search, and the Gaussian primitives are divided into load-bearing chains and non-load-bearing regions accordingly. In step one, the method for reconstructing the cutting force, vibration, and temperature signals into a continuous physical field using neural operators is as follows: Fourier neural operator (FNO) or deep operator network (DeepONet) is used to learn the operator mapping from the continuous physical field observed by sparse sensing to reconstruct the temperature field, vibration field, and cutting force field; and a graph neural network (GNN) is used to pass messages along the assembly topology on the load-bearing chain, so that the vibration and cutting force propagate along the kinematic chain and an impedance attenuation factor is introduced across the assembly interface to mask the non-load-bearing area. In step one, global or regional scalar parameters characterizing the processing environment state are also collected, and a global environment state vector is constructed. The global environment state vector This includes one or more of the following: relative humidity, air pressure, low-frequency environmental vibration, cutting fluid flow rate, cutting fluid pressure, and pneumatic system pressure; the global environmental state vector. Instead of anchoring each primitive, the hidden state encoding is incorporated in step two and used as a conditional input to the forward dynamic world model in step three. In step one, the appearance mode, temperature mode, vibration mode and cutting force mode of the physically perceived three-dimensional Gaussian sputtering twin field share the same rasterization rendering pipeline. When switching modes, only the attribute buffer pointer is rebound without rebuilding the spatial acceleration structure.

4. The method for constructing and autonomously deciding on the embodied world model of a precision CNC machine tool according to claim 1, characterized in that: In step two, the system state It also includes a gradually changing hidden state characterizing tool wear or breakage, as well as the initial stress or surface hardness of the workpiece; the gradually changing hidden state is estimated online by cutting force signal characteristics, acoustic emission signals or visual inspection results, and continuously corrected by the residual in step six, serving as the conditional input of the forward dynamic world model; In step two, processing-related information is hierarchically carried according to its physical properties. The method is as follows: physical fields with spatial distribution characteristics are carried in the physically-aware three-dimensional Gaussian sputtering twin field in a per-element anchoring manner; global or regional scalars without spatial distribution characteristics are carried in the global environment state vector. The tool and workpiece states, which are carried and used as conditions in the forward dynamic world model, vary gradually with the machining process and are carried and estimated online in the gradually changing hidden states.

5. The method for constructing and autonomously deciding on the embodied world model of a precision CNC machine tool according to claim 1, characterized in that: In step three, the forward dynamic world model is composed of one or more of the following: the cyclic state-space model RSSM, the diffusion transformer operating in the latent space, or the joint embedded prediction architecture JEPA; the forward dynamic world model is trained with historical interaction data, which is composed of parametric simulation data and actual machine processing data, and the gap between simulation and reality is bridged by domain randomization. The training objective of the forward dynamic world model includes a physical consistency regularization term, which constrains the predicted state to satisfy physical priors such as cross-interface impedance attenuation, propagation along the kinematic chain, time-scale separation, and causal consistency between the environmental state and the processing state in the form of the residual of a neural operator or a physical information neural network PINN.

6. The method for constructing and autonomously deciding on the embodied world model of a precision CNC machine tool according to claim 1, characterized in that: In step four, the candidate processing action sequence is sampled and extrapolated in parallel using Model Predicted Path Integral (MPPI), Cross-Entropy Method (CEM), or diffusion strategy. The multi-objective cost function is a weighted sum of predicted surface roughness, predicted vibration amplitude, predicted geometric error, energy consumption and time penalty terms, and environmental robustness penalty terms. The environmental robustness penalty term increases the cost when the extrapolated environmental state or extended physical field in the next few steps exceeds a preset safety threshold. Step four is repeated in a rolling time domain manner to achieve rolling replanning. The forward dynamic world model is a causal world model that intervenes in the causal relationship and uses counterfactual imagination on candidate processing actions, and enables the imaginative autonomous decision-making planning to prioritize robust processing action sequences that are insensitive to environmental disturbances and operating condition drift.

7. The method for constructing and autonomously deciding on the embodied world model of a precision CNC machine tool according to claim 1, characterized in that: In step five, the execution kernel performs finite impulse response (FIR) exponential decay low-pass filtering on the acquired raw signal, and uses the root mean square amplitude of the filtered vibration and its rate of change to form a state vector; under the guidance of the action target given by the cognitive kernel, a multi-variable collaborative compensation quantity is generated by the compensation law subject to saturation and rate constraints. The multi-variable collaborative compensation quantity is the feed rate and spindle speed, or two or more of the following: grinding wheel linear speed, workpiece speed, axial feed and grinding depth. The high-frequency channel is responsible for rapid suppression, and the low-frequency channel is responsible for depth adjustment.

8. A system for constructing and autonomously deciding on a world model of a precision CNC machine tool, used to implement the method for constructing and autonomously deciding on a world model of a precision CNC machine tool as described in any one of claims 1-7, characterized in that: include: The multimodal embodied perception module is used to collect cutting force, vibration, temperature signals, multi-view images, and environmental perception signals characterizing the state of the machining environment during the machine tool machining process. PA-3DGS twin construction module is used to reconstruct and build a physical perception 3D Gaussian sputtering twin field that is injected with CAD component-level semantics, reconstructed by neural operators, and anchored to the physical perception channel on a per-primary basis. The embedded hidden state encoding module is used to compress the twin field, multimodal signals and the environmental perception signals into a system state, and incorporate the tool and workpiece states that change slowly with the machining process as slowly changing hidden states into the system state. The forward dynamic world model module under action conditions is used to imagine and predict the future system state, surface roughness, vibration amplitude and geometric error based on candidate processing actions and the environmental perception signals. The imaginative autonomous decision-making and planning module is used to deduce and select the best processing action from the candidate processing action sequence and output the optimal processing action; The dual-timescale asynchronous dual-core execution module includes an asynchronous parallel cognitive core and an execution core that share the system state, used to complete imaginary planning and constrained servo compensation delivery under the condition of satisfying singular perturbation stability. The embodied closed-loop online continuous self-learning module is used to fine-tune the world model, update the slowly varying hidden states, and back-anchor the prediction results to the rendering by continuously learning online using prediction-measurement residuals.

9. The system for constructing and autonomously deciding on the embodied world model of a precision CNC machine tool according to claim 8, characterized in that: It also includes a vision-language-action (VLA) task interface, which uses natural language or task-level instructions to set the weights of the processing objectives and the multi-objective cost function, and is then resolved into planning objectives by the embodied world model.

10. An embodied intelligence system, characterized in that: It includes multiple precision CNC machine tool embodied world model construction and autonomous decision-making systems as described in any one of claims 8-9. Each precision CNC machine tool embodied world model construction and autonomous decision-making system serves as a machine tool-level single machine body. The machine tool-level single machine body is subjected to multi-machine collaborative imagination and scheduling through the workshop-level embodied world model, and bidirectional interaction is carried out through federated continuous learning and knowledge distillation with the basic model layer of the cloud manufacturing world model to realize collaborative forward-looking inference and autonomous decision-making of the coupled state of workpiece, machine tool and logistics.

Images