Cutter parameter dynamic matching method and device based on deep learning

Abstract

The invention provides a tool parameter dynamic matching method and device based on deep learning, and relates to the field of data processing. The method comprises the following steps: acquiring laser holographic imaging data and three-dimensional contour scanning data of a machine tool cutting area at a target timestamp; according to the laser holographic imaging data and the three-dimensional contour scanning data, holographic voxel data and cutting point cloud data are determined; inputting the holographic voxel data and the clipping point cloud data into a four-dimensional grid containing a target timestamp to obtain a four-dimensional snapshot sequence; acquiring a first cutter parameter of the machine tool cutter at the target timestamp; and inputting the four-dimensional snapshot sequence and the first cutter parameter into a preset deep learning causal hypernet to generate a second cutter parameter. By implementing the technical scheme provided by the invention, the cutter parameters can be dynamically matched during machining of the numerical control machine tool.

Description

Technical Field

[0001] The present application relates to the technical field of data processing, and specifically to a method and device for dynamic matching of tool parameters based on deep learning. Background Art

[0002] As the manufacturing industry develops towards intelligence, flexibility and efficiency, the adjustment of CNC machine tool tool parameters has become a key link in improving processing quality and production efficiency.

[0003] Currently, existing technologies often use traditional vibration sensing and two-dimensional vision systems to adjust tool parameters. However, due to the inherent sampling principles and resolution limitations of this approach, it is difficult to capture microscopic topographical changes in the contact area between the tool and the workpiece. For example, vibration signals can only reflect changes in overall cutting forces and machine tool stiffness, and are almost insensitive to micron-level tool wear or surface details. Two-dimensional cameras, affected by viewing angle, lighting, and occlusion, can only provide projection images of the tool side or the workpiece side, and cannot reproduce the dynamic evolution of three-dimensional microstructures. Therefore, these methods are not conducive to dynamically matching tool parameters during CNC machine tool processing.

[0004] Therefore, there is an urgent need for a tool parameter dynamic matching method and device based on deep learning. Summary of the Invention

[0005] The present application provides a method and device for dynamic matching of tool parameters based on deep learning, which facilitates dynamic matching of tool parameters during CNC machine tool processing.

[0006] In a first aspect of the present application, a tool parameter dynamic matching method based on deep learning is provided, the method comprising: acquiring laser holographic imaging data and three-dimensional contour scanning data for a machine tool cutting area at a target timestamp; determining holographic voxel data and cropped point cloud data based on the laser holographic imaging data and the three-dimensional contour scanning data; inputting the holographic voxel data and the cropped point cloud data into a four-dimensional grid containing the target timestamp to obtain a four-dimensional snapshot sequence; acquiring a first tool parameter of a machine tool at the target timestamp; inputting the four-dimensional snapshot sequence and the first tool parameter into a preset deep learning causal supernet to generate a second tool parameter.

[0007] By employing the aforementioned technical solution, a multimodal sensing approach combining laser holographic imaging and 3D contour scanning, combined with high-resolution voxelization and point cloud cropping, is constructed to accurately reflect the spatiotemporal changes in the cutting zone. This not only significantly enhances the visualization and quantification of tool-workpiece interaction, but also enables a leap in cutting state information from 2D projections to 3D structures and even 4D dynamics. Furthermore, by integrating these 4D snapshots with current tool parameters into a pre-defined causal hypernet model, this method overcomes the bottlenecks of traditional vibration sensing and 2D vision systems, which rely heavily on correlation and lack interpretability. This approach not only enables reasonable predictions based on historical states but also explores the deep logical relationship between tool parameter adjustments and actual cutting results from a causal-driven perspective, thereby generating more physically meaningful and reliable secondary tool parameters. The overall solution offers the advantages of high precision, adaptability, and interpretability, significantly improving the responsiveness and intelligent level of tool parameter adjustment, and providing strong technical support for adaptive cutting and intelligent decision-making under complex working conditions. This facilitates dynamic matching of tool parameters during CNC machine tool machining.

[0008] Optionally, the acquiring of laser holographic imaging data and three-dimensional contour scanning data for the machine tool cutting area at the target timestamp specifically includes: sending synchronous acquisition instructions to the laser holographic camera and the three-dimensional contour scanner respectively; reading the timestamps returned by the laser holographic camera and the three-dimensional contour scanner, and determining that the timestamp is the target timestamp; under the constraint of the target timestamp, receiving the interference fringe sequence sent by the laser holographic camera, and receiving the point cloud batch sent by the three-dimensional contour scanner, to obtain the laser holographic imaging data and the three-dimensional contour scanning data.

[0009] By adopting the above technical solution, precise synchronous acquisition control is achieved between the laser holographic camera and the 3D profile scanner, ensuring that both complete data acquisition at the same target timestamp. This effectively solves the time misalignment and information inconsistency problems that exist in traditional multimodal sensors in processing monitoring. Specifically, by actively sending synchronous acquisition instructions and reading the return timestamp, the interference fringe sequence and point cloud batch are strictly aligned in time, making the two types of data highly timely and corresponding. This design not only avoids the complex subsequent timing calibration process and reduces the risk of feature mismatch caused by time drift, but also lays a high-quality data foundation for the subsequent construction of a unified four-dimensional snapshot. Furthermore, the laser holographic camera can capture microscopic interference information, reflecting implicit structural characteristics such as internal stress and deformation of the material, while the 3D profile scanner provides macroscopic morphology and surface contour changes. Through the coordinated fusion of dual-source data, dynamic perception from microscopic stress evolution to macroscopic morphological changes can be achieved, greatly improving the comprehensiveness and depth of cutting process modeling.

[0010] Optionally, determining the holographic voxel data and the cropped point cloud data based on the laser holographic imaging data and the three-dimensional contour scanning data specifically includes: performing de-artifacting and brightness equalization on the interference fringe pattern contained in the laser holographic imaging data to obtain an initialization image; mapping the initialization image into three-dimensional voxels according to a predefined spatial grid, each of the three-dimensional voxels including a corresponding local phase and local light intensity; performing statistical filtering on the contour scanning point cloud contained in the three-dimensional contour scanning data to obtain an initialization point cloud; determining the surface normal on the initialization point cloud to obtain an intended area; and converting the three-dimensional voxel and the intended area into a unified coordinate and data structure, respectively, to obtain the holographic voxel data and the cropped point cloud data.

[0011] By employing this technical solution, artifact removal and brightness equalization are first performed on the interference fringe pattern in the laser holographic data, effectively suppressing interference from factors such as environmental noise and uneven illumination, significantly improving image quality and physical measurement accuracy. Subsequently, by mapping the optimized image onto a predefined 3D grid structure, a voxel field containing local phase and intensity information is constructed. This not only spatially encodes microstructural changes but also preserves the continuity and interpretability of the material stress field and deformation process. Furthermore, statistical filtering is used to remove outliers and outliers in the 3D contour scan data, improving the density and validity of the point cloud. Surface normals are further calculated to determine the intended machining area, allowing the processing process to focus more on the actual tool-workpiece interaction, enhancing data utilization efficiency and the relevance of downstream models. Finally, the holographic voxels and the cropped point cloud are transformed into unified coordinates and structures, enabling the fusion of heterogeneous data at the spatial and semantic levels, significantly reducing the complexity of multimodal information in joint modeling.

[0012] Optionally, inputting the holographic voxel data and the cropped point cloud data into a four-dimensional grid containing the target timestamp to obtain a four-dimensional snapshot sequence specifically includes: obtaining the engineering requirements of the machine tool cutting area; setting the spatial three-dimensional resolution and time step according to the engineering requirements; filling the holographic voxel data into the four-dimensional grid according to the three-dimensional resolution to obtain a first moment slice; filling the cropped point cloud data into the four-dimensional grid according to the time step to obtain a second moment slice; and splicing the first moment slice and the second moment slice in chronological order to obtain the four-dimensional snapshot sequence.

[0013] By adopting the above technical solution, by sequentially embedding the holographic voxel data and the cropped point cloud data into a four-dimensional grid structure containing time information, not only is the deep fusion of multi-source data achieved in a unified spatiotemporal framework, but the continuity and expressiveness of the dynamic modeling of the cutting process are also significantly improved. First, by obtaining the engineering requirements of the machine tool cutting area and setting the appropriate spatial three-dimensional resolution and time step accordingly, it is ensured that the constructed four-dimensional grid can take into account both data accuracy and real-time performance, satisfying the detailed expression of the evolution of the micro-morphology while ensuring the reasonable allocation of computing resources and the response speed of the model. In the spatial dimension, the holographic voxel data reflects the structural characteristics such as the stress distribution and material deformation inside the cutting area. By mapping it to the first moment slice in the grid, the initial state information of the processing can be finely encoded; in the temporal dimension, the cropped point cloud data dynamically reflects the trajectory of the surface morphology evolution over time. By filling in the subsequent moment slices, the evolution trend of the micro-geometric structure during the cutting process is truly reproduced. Finally, slices from different time points were spliced ​​together in chronological order to construct a complete four-dimensional snapshot sequence, successfully achieving the dimensionality-upgrading expression of the evolution process from static voxel and point cloud data to dynamic, continuous space.

[0014] Optionally, obtaining the first tool parameters of the machine tool tool at the target timestamp specifically includes: querying the cutting speed, feed rate and cutting depth of the machine tool tool at the target timestamp from the CNC controller; synchronously obtaining the tool model and wear level of the machine tool tool; determining the first tool parameters based on the cutting speed, feed rate, cutting depth, tool model and wear level.

[0015] By adopting the above technical solutions, firstly, cutting speed, feed rate and cutting depth as the core process parameters that determine machining efficiency, surface quality and tool load directly reflect the operating behavior of the tool under the current machining conditions. Accurately extracting these indicators can effectively lock the actual machining state and avoid prediction deviations caused by inconsistencies between offline settings and online reality. Secondly, tool model and wear level as key descriptions reflecting the physical properties of the tool cover the combined influence of tool geometry, material composition and its service life stage, and can reveal changes in cutting performance due to differences in tool health status under the same process parameters. By synchronously acquiring and jointly modeling these five types of parameters, the first tool parameter formed not only has the operation instruction information of the process dimension, but also includes the state description of the physical level, realizing the precise association between operation, carrier and performance.

[0016] Optionally, the four-dimensional snapshot sequence and the first tool parameters are input into a preset deep learning causal supernet to generate the second tool parameters, specifically including: normalizing and dimensionally reducing the four-dimensional snapshot sequence and the first tool parameters to obtain a spatiotemporal feature summary and tool parameter features; concatenating the spatiotemporal feature summary and the tool parameter features into a unified vector, and appending key node indicators marked in a preset causal graph to obtain a target analysis vector, wherein the key node indicators include material type and ambient temperature; using the preset deep learning causal supernet constructed by the preset causal graph, inferring the target analysis vector in combination with the supernet weights to obtain an inference parameter adjustment amount; and superimposing the inference parameter adjustment amount and the first tool parameters to obtain the second tool parameters.

[0017] By adopting the above technical solution, a causal supernet inference framework is introduced to process the four-dimensional snapshot sequence and the first tool parameters. This overcomes the limitations of traditional data, such as poor interpretability and weak generalization, and significantly improves the scientific and intelligent level of tool parameter adjustment. First, feature normalization and dimensionality reduction are used to process the four-dimensional snapshot sequence and tool parameters, effectively compressing high-dimensional redundant information while retaining key spatiotemporal trends and parameter variation characteristics. This ensures that the input vector maintains the stability of model training and enhances semantic expression. Subsequently, by concatenating the extracted spatiotemporal feature summary with the tool parameter features into a unified vector and incorporating key node indicators such as material type and ambient temperature, a physical logic-based target analysis vector is constructed. This structure fully reflects the impact path of multi-factor interactions on cutting behavior, expands the model's perceptual boundaries, and enables the inference process to move beyond superficial statistical correlations and integrate causal constraints from the real-world machining context. The inference results are output as parameter adjustment values, which are weighted and superimposed with the first tool parameters to generate the second tool parameters. This ensures that the adjustment results are not only highly responsive, but also dynamically adaptable and practically feasible.

[0018] Optionally, the method further includes: receiving user-defined tool parameters sent by a user for the machine tool cutting area; and controlling the machine tool tool to dynamically match the user-defined tool parameters according to the user-defined tool parameters.

[0019] By adopting the above technical solution, by introducing user-defined tool parameters and applying them directly to machine tool tools as dynamic control instructions, not only a closed-loop link of human-machine interaction, intelligent decision-making, and real-time control is established, but also the personalized response capability and flexible manufacturing level of the entire system in complex processing environments are significantly enhanced. First, users can independently set tool parameters, including but not limited to key indicators such as cutting speed, feed rate, and cutting depth, based on specific process requirements, processing tasks, or real-time observation results. This hybrid decision-making mechanism of human intervention + data-driven breaks the passive execution control mode in traditional automation systems and forms an organic synergy between human experience and intelligent reasoning systems. It is particularly suitable for dealing with complex scenarios such as sudden interference during processing, personalized customization, or uncertain material properties.

[0020] In a second aspect of the present application, a tool parameter dynamic matching device based on deep learning is provided, the device comprising an acquisition module and a processing module, wherein the acquisition module is used to acquire laser holographic imaging data and three-dimensional contour scanning data for a machine tool cutting area at a target timestamp; the processing module is used to determine holographic voxel data and cropped point cloud data based on the laser holographic imaging data and the three-dimensional contour scanning data; the processing module is further used to input the holographic voxel data and the cropped point cloud data into a four-dimensional grid containing the target timestamp to obtain a four-dimensional snapshot sequence; the acquisition module is further used to acquire a first tool parameter of a machine tool tool at the target timestamp; the processing module is further used to input the four-dimensional snapshot sequence and the first tool parameter into a preset deep learning causal supernet to generate a second tool parameter.

[0021] In a third aspect of the present application, an electronic device is provided, which includes a processor, a memory, a user interface, and a network interface. The memory is used to store instructions, the user interface and the network interface are both used to communicate with other devices, and the processor is used to execute the instructions stored in the memory so that the electronic device performs the method described above.

[0022] In a fourth aspect of the present application, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores instructions. When the instructions are executed, the method described above is executed.

[0023] In summary, one or more technical solutions provided in this application have at least the following technical effects or advantages: By incorporating multimodal sensing techniques from laser holography and 3D contour scanning, combined with high-resolution voxelization and point cloud cropping, a four-dimensional snapshot sequence accurately reflects the spatiotemporal changes in the cutting zone. This not only significantly enhances the visualization and quantification of tool-workpiece interaction, but also enables a leap in cutting state information from two-dimensional projections to three-dimensional structures and even four-dimensional dynamics. Furthermore, by integrating these four-dimensional snapshots with current tool parameters into a pre-defined causal hypernet model, this approach overcomes the bottlenecks of traditional vibration sensing and two-dimensional vision systems, which rely heavily on correlation and lack interpretability. This approach not only enables reasonable predictions based on historical states but also explores the deep logical relationship between tool parameter adjustments and actual cutting results from a causal-driven perspective, thereby generating more physically meaningful and reliable secondary tool parameters. This overall solution offers the advantages of high precision, adaptability, and interpretability, significantly improving the responsiveness and intelligent level of tool parameter adjustment, and providing strong technical support for adaptive cutting and intelligent decision-making under complex working conditions. This facilitates dynamic matching of tool parameters during CNC machine tool machining. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 A schematic flow chart of a tool parameter dynamic matching device method based on deep learning provided in an embodiment of the present application; Figure 2 Another flowchart of a tool parameter dynamic matching device method based on deep learning provided in an embodiment of the present application; Figure 3 A schematic diagram of a module of a tool parameter dynamic matching device based on deep learning provided in an embodiment of the present application; Figure 4 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application.

[0025] Explanation of the reference numerals: 31, acquisition module; 32, processing module; 41, processor; 42, communication bus; 43, user interface; 44, network interface; 45, memory. DETAILED DESCRIPTION

[0026] In order to enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the drawings in the embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this application, not all of the embodiments.

[0027] In the description of the embodiments of this application, words such as "for example" or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as "for example" or "for instance" in the embodiments of this application should not be construed as being preferred or advantageous over other embodiments or designs. Rather, the use of words such as "for example" or "for instance" is intended to present the relevant concepts in a concrete manner.

[0028] In the description of the embodiments of the present application, the term "multiple" means two or more. For example, multiple systems refer to two or more systems, and multiple screen terminals refer to two or more screen terminals. In addition, the terms "first" and "second" are used for descriptive purposes only and are not to be understood as indicating or implying relative importance or implicitly indicating the indicated technical features. Thus, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. The terms "including", "comprising", "having" and their variations all mean "including but not limited to", unless otherwise specifically emphasized.

[0029] As the manufacturing industry continues to move towards intelligent, flexible and efficient development, the dynamic adjustment of tool parameters in CNC machine tools has gradually become one of the core technologies for improving processing quality and production efficiency.

[0030] However, existing technologies mostly rely on traditional vibration sensors or two-dimensional vision systems to achieve tool status perception and parameter adjustment. Due to their own perception mechanisms and physical resolution capabilities, they are unable to meet the real-time grasp of microscopic information required in high-precision machining scenarios.

[0031] Specifically, vibration sensors primarily respond to the macroscopic mechanical characteristics of the overall cutting process, such as fluctuations in cutting forces and changes in structural rigidity, and have little ability to perceive key details such as micron-level tool wear, cutting edge defects, or micro-deformations of the contact surface. Two-dimensional vision systems, however, are limited by single-view observation, changing lighting conditions, and occlusion. The images they capture are merely two-dimensional projections of the tool or workpiece surface, unable to reconstruct the true three-dimensional topography, let alone capture its evolution during dynamic machining. For this reason, current methods often suffer from parameter adjustment lags, inaccurate responses, and insufficient basis for adjustment when dealing with highly complex, high-precision machining tasks, making it difficult to support intelligent, dynamic matching of machine tool tool parameters.

[0032] In order to solve the above technical problems, this application provides a tool parameter dynamic matching method based on deep learning, referring to Figure 1 , Figure 1 This is a flowchart of a method for dynamic tool parameter matching based on deep learning provided in an embodiment of the present application. The method is applied to a server and includes steps S110 to S150, which are as follows: S110 , acquiring laser holographic imaging data and three-dimensional contour scanning data of a machine tool cutting area at a target time stamp.

[0033] Specifically, a back-end server in the system is responsible for sending data read requests to two types of high-precision sensors placed in the cutting area of ​​the CNC machine tool at a precise moment (i.e., the target timestamp), and saving the original detection information they collect for subsequent analysis and processing. Specifically, laser holographic imaging data is a sequence of fringe images captured by laser interferometer cameras placed around the contact area between the tool and the workpiece. These interference fringes contain stress distribution and morphological changes at the micron or even submicron level in the processing area; three-dimensional contour scanning data is a collection of point clouds from high-speed laser profiling or confocal scanners, which are used to reconstruct the geometric contours of the cutting surface and workpiece surface. The server uses the target timestamp as a time reference to ensure that the two types of data can be accurately aligned at the same time point, thereby eliminating errors caused by clock drift or trigger delays in different devices.

[0034] For example, in an aviation parts processing workshop, when the machine tool reaches the critical moment of 15.234 seconds into a certain cut, the server sends a synchronous acquisition command to the laser holographic camera, generating a set of interference fringe patterns representing that moment. Simultaneously, it initiates a point cloud sampling request to the 3D contour scanner, obtaining tens of thousands to millions of 3D coordinate points at that moment. After receiving these two sets of data, the server stores them in a database with a unified time tag, labeling them with something like "Cutting Zone - 15.234s." Subsequently, both digital twin-based 4D field reconstruction and causal hypernet parameter prediction can accurately call upon these two aligned high-precision sensor data sets, ensuring the spatiotemporal consistency and physical authenticity of the model inputs.

[0035] In one possible implementation, laser holographic imaging data and three-dimensional contour scanning data for a machine tool cutting area at a target timestamp are obtained, specifically including: sending synchronization acquisition instructions to the laser holographic camera and the three-dimensional contour scanner, respectively; reading the timestamps returned by the laser holographic camera and the three-dimensional contour scanner, and determining that the timestamp is a target timestamp; under the constraint of the target timestamp, receiving the interference fringe sequence sent by the laser holographic camera and the point cloud batch sent by the three-dimensional contour scanner, to obtain the laser holographic imaging data and the three-dimensional contour scanning data.

[0036] Specifically, first, the server sends a "start acquisition" instruction to each of the two devices. This instruction carries a pre-set target timestamp, telling the sensor to start data capture at that moment. Next, the sensor will complete the data capture internally, and stamp each frame of the interference fringe pattern or each batch of point cloud data with a local timestamp and send it back to the server. After receiving the data, the server will compare the timestamp returned by the sensor with the predetermined target timestamp to see if it matches. Only when the two are completely consistent or within an acceptable range of small time errors will the server officially receive and archive the collected data. Otherwise, it will retry or alarm. This closed-loop mechanism of "pre-triggering, local acquisition, time verification, and confirmed storage" eliminates the risk of asynchrony caused by device response delays, network jitter, or internal clock drift to the greatest extent, ensuring that the obtained holographic image and three-dimensional point cloud are accurately aligned in time, laying a solid foundation for subsequent four-dimensional reconstruction and causal analysis.

[0037] For example, in an aerospace parts shop, to monitor tool microwear during machining of titanium alloy blades, a server sends acquisition commands via Industrial Ethernet to a laser holographic camera and a high-speed 3D profile scanner installed in the cutting area at the critical moment of 12.345 seconds after machining begins. The holographic camera immediately records a high-resolution image containing hundreds of interference fringes and marks it with the time stamp of 12.345 seconds. Simultaneously, the profile scanner captures approximately two million 3D surface points, also annotated with the same timestamp. After comparing the two data sets and confirming that the timestamps match, the server archives them as "BladeCut_12.345s_holo" and "BladeCut_12.345s_cloud" for subsequent 3D structure reconstruction and dynamic topography analysis at the same time. This alignment accuracy reaches milliseconds or even microseconds, ensuring that even the smallest changes during the machining process are not missed, truly enabling highly precise and synchronized tool and workpiece condition monitoring.

[0038] S120 , determining holographic voxel data and cropped point cloud data according to the laser holographic imaging data and the three-dimensional contour scanning data.

[0039] Specifically, the server first reads the denoised and brightness-corrected laser holographic interference fringe pattern and divides it into a number of small cubic units (voxels) according to a predefined three-dimensional spatial grid. Within each voxel, the server calculates the phase difference and light intensity distribution of the corresponding pixel in that spatial unit, thereby generating a set of "holographic voxel data." These voxels not only preserve the continuous spatial information of material stress and micro-deformation within the cutting zone but also unify the coordinate system of different sensing frames, facilitating subsequent spatiotemporal data fusion. For example, during a titanium alloy machining process, the server maps the 512×512 pixel interference pattern onto a 200×200×200 cubic voxel grid. Each voxel records the local phase peak and light intensity mean, perfectly reproducing the submicron-level stress distribution at the tool-workpiece interface.

[0040] At the same time, the server statistically filters the raw point cloud data provided by the 3D contour scanner to remove isolated noise points and invalid reflection points. It then automatically identifies the "intended area"—the portion of the point cloud where the tool actually contacts the workpiece—based on the predicted tool geometry or the surface normals calculated in real time. The server then retains only the set of valid points within this area, forming the "cropped point cloud data" and converting it to the same coordinate system as the holographic voxels. For example, when the scanner captures 500,000 scattered points, a triple standard deviation filter leaves approximately 450,000 valid points. These points are then cropped to within 2 mm of the tool contour, extracting only approximately 60,000 key points. This ensures that subsequent spatiotemporal modeling focuses on key information while avoiding irrelevant background interference. The resulting holographic voxels and cropped point cloud are both efficient and accurate, providing high-quality structured data support for constructing four-dimensional snapshots and causal reasoning.

[0041] In one possible implementation, holographic voxel data and cropped point cloud data are determined based on laser holographic imaging data and three-dimensional contour scanning data, specifically including: performing de-artifacting and brightness equalization on the interference fringe pattern contained in the laser holographic imaging data to obtain an initialization image; mapping the initialization image into three-dimensional voxels according to a predefined spatial grid, each three-dimensional voxel including a corresponding local phase and local light intensity; performing statistical filtering on the contour scanning point cloud contained in the three-dimensional contour scanning data to obtain an initialization point cloud; determining the surface normal on the initialization point cloud to obtain an intended area; and converting the three-dimensional voxel and the intended area into unified coordinates and data structures, respectively, to obtain holographic voxel data and cropped point cloud data.

[0042] Specifically, the server first performs artifact removal and brightness balancing on the interference fringe pattern captured by the laser holographic camera. For example, when factors such as mirror reflections or dust create abrupt dark spots or highlights in the image, the artifact removal algorithm automatically identifies and repairs these local distortions, while brightness balancing ensures uniform light intensity distribution across the entire image, preventing loss of detail due to dark edges or bright centers. After processing, the system converts each two-dimensional interference pattern "slice" into a series of voxels based on a pre-defined three-dimensional spatial grid (for example, a 200×200×200 cube). Each voxel contains both the phase information of that subregion (reflecting minor deformations) and the light intensity value (reflecting changes in material density or refractive index), accurately reproducing the microscopic stress and topographical characteristics of the cut area in three dimensions.

[0043] In parallel, the server performs statistical filtering and normal estimation on the raw point cloud output by the 3D contour scanner to extract the truly valuable "intent region." For example, a scanner may capture nearly a million raw points at a time, including debris and stray reflections from the machine base, fixtures, and the tool-workpiece contact area. Statistical filtering removes noise points that are isolated from the main cluster (such as flying chips), while normal calculation helps identify the smooth area where the tool edge actually contacts the workpiece surface. The system retains only these local point clouds, such as 50,000 valid points within 0-2 mm from the tool tip, and maps them to the same coordinate system and data structure as the holographic voxels. The resulting "holographic voxel data" and "cropped point cloud data" are two highly aligned, uniformly formatted 3D descriptions. They preserve the hologram's detailed depiction of the internal stress field while highlighting the point cloud's geometric restoration of the external contour, providing high-quality, directly callable input for subsequent 4D mesh construction and causal hypernet inference.

[0044] S130 , inputting the holographic voxel data and the cropped point cloud data into a four-dimensional grid containing a target timestamp to obtain a four-dimensional snapshot sequence.

[0045] Specifically, the server will first create a four-dimensional array in memory based on the preset spatial resolution (for example, one grid per cubic millimeter) and time step (for example, one frame per millisecond), with its coordinate axes corresponding to the three spatial dimensions X, Y, and Z and the time dimension T. It will then map the holographic voxel data at the target timestamp onto the three-dimensional slice at that time point, with each grid unit recording the phase and light intensity information of the corresponding position. Similarly, the cropped point cloud data at the same timestamp is projected onto the three-dimensional slice at the same position, and the voxel information is supplemented with geometric point and normal attributes. In this way, a single "time point" is no longer just a plane or a set of points, but a complete three-dimensional body that mixes microscopic internal structure with external geometric contours.

[0046] For example, suppose the server is machining an aviation aluminum alloy plate with a spatial resolution of 0.5 mm and a time step of 2 milliseconds. At frame t = 10.246s, it first fills the 200×200×100 voxels obtained from the holographic imaging at that moment into a grid slice, reconstructing a volumetric map of the internal stress distribution. It then maps the approximately 50,000 cropped point clouds acquired simultaneously onto the same slice, highlighting the surface microcontours and wear patterns. The server then repeats the same operation at t = 10.248s, 10.250s, and so on, in 2-millisecond increments, ultimately outputting a sequence of dozens or even hundreds of consecutive four-dimensional snapshots. This sequence not only fully reproduces the evolution of the tool-workpiece contact zone at the microscopic level but also provides high-fidelity, continuous, and consistent input data for subsequent dynamic analysis and parameter optimization based on neural fields or causal hypernets.

[0047] In one possible implementation, the holographic voxel data and the cropped point cloud data are input into a four-dimensional grid containing a target timestamp to obtain a four-dimensional snapshot sequence, specifically including: obtaining the engineering requirements of the machine tool cutting area; setting the spatial three-dimensional resolution and time step according to the engineering requirements; filling the holographic voxel data into the four-dimensional grid according to the three-dimensional resolution to obtain a first moment slice; filling the cropped point cloud data into the four-dimensional grid according to the time step to obtain a second moment slice; and splicing the first moment slice and the second moment slice in chronological order to obtain a four-dimensional snapshot sequence.

[0048] Specifically, the system first obtains the processing objectives from the process engineer—for example, how many millimeters of micro-deformation need to be monitored, and how high a temporal resolution is required to capture rapid vibration or wear processes. Then, based on these requirements, the grid size in each spatial dimension (e.g., per cubic millimeter, per cubic ten microns) and the sampling interval in the temporal dimension (e.g., every 2 milliseconds, every 5 milliseconds) are set. This customized resolution and step size configuration ensures that the grid can accurately reproduce the subtle differences in stress and topography within the cutting area, while also avoiding the data redundancy and computational bottlenecks caused by excessively high resolution, leaving a controllable performance space for subsequent analysis and model reasoning.

[0049] Next, the system will sequentially fill the previously obtained holographic voxel data and cropped point cloud data into different "time slices" of the four-dimensional grid according to the above settings: at a critical moment (such as the moment the tool begins to cut into the workpiece), the voxelized holographic information is first mapped to the first time slice to fully express the microscopic phase and light intensity distribution; then, at a predetermined time step (such as 3 milliseconds), the updated point cloud geometry information is filled into the next slice to reflect the slight changes in the surface contour. By connecting these three-dimensional slices end to end in chronological order, the system ultimately generates a continuous sequence of four-dimensional snapshots - like piecing together a series of 3D snapshots to form a "movie" of the cutting area's microscopic evolution over time. For example, in aviation parts processing, engineers may require a spatial resolution of 0.5 mm and a time step of 2 milliseconds. In this case, the system will fill in a frame of holographic voxels at t=10.000s, a frame of cropped point cloud at t=10.002s, and the next frame of voxels at t=10.004s. This cycle will continue to expand to form a complete four-dimensional sequence, providing a highly consistent, continuous and smooth spatiotemporal data foundation for refined modeling, causal analysis and online tool parameter optimization.

[0050] S140 , obtaining the first tool parameter of the machine tool at the target timestamp.

[0051] Specifically, the server first reads the three core parameters corresponding to that point in time: cutting speed, feed rate, and depth of cut. It then retrieves the model identification of the currently installed tool and the real-time wear level or tool life indicator recorded by the system. This allows the server to accurately identify the tool's fundamental process parameters at that moment, including cutting speed, material feed rate, and depth of cut, while also taking into account the tool's physical state. This provides a complete and consistent initial basis for subsequent model input and parameter tuning.

[0052] In one possible implementation, obtaining the first tool parameters of a machine tool tool at a target timestamp specifically includes: querying the cutting speed, feed rate, and cutting depth corresponding to the machine tool tool at the target timestamp from a CNC controller; synchronously obtaining the tool model and wear level of the machine tool tool; and determining the first tool parameters based on the cutting speed, feed rate, cutting depth, tool model, and wear level.

[0053] Specifically, the server first sends a real-time request to the CNC controller via an industrial communication protocol (such as OPC-UA or MTConnect) to accurately obtain the three core process parameters at that moment: cutting speed, feed rate, and depth of cut. These three parameters directly reflect the tool's motion and cutting conditions under the current machining load and serve as essential data for evaluating machining efficiency and surface quality. The server then simultaneously queries the tool management system or the tool magazine maintained internally by the CNC to retrieve the model identification (including geometry, coating type, etc.) of the tool currently in use, as well as the wear level (e.g., new tool, light wear, moderate wear, etc.) determined through online monitoring or pre-defined maintenance systems. The entire process requires time alignment and data consistency to ensure that each acquired parameter precisely matches the target timestamp without delay or frame dropouts, providing reliable and synchronized input for subsequent modeling and adjustments.

[0054] For example, when machining a batch of aerospace aluminum alloy parts, at the critical juncture of 17.892 seconds into the machining program, the server automatically initiates a data request to the CNC controller, receiving a cutting speed of 140 m / min, a feed rate of 0.18 mm / rev, and a depth of cut of 2.8 mm. Simultaneously, through the same communication session, it also retrieves the corresponding tool model "TiAlN-Coated Ø8mm EndMill" from the tool magazine and its "mild wear" (approximately 25% of service life) status as determined by the online monitoring system. The server packages these process parameters and tool status into a complete set of "first tool parameters," encompassing both the tool's current motion and cutting conditions and its own physical health. This set of parameters can be directly fed back to the monitoring interface for engineers to review or fed into intelligent algorithms such as causal hypernets to generate more accurate and safer second tool parameter adjustment plans.

[0055] S150 , inputting the four-dimensional snapshot sequence and the first tool parameters into a preset deep learning causal supernet to generate second tool parameters.

[0056] Specifically, the server first aggregates a formatted sequence of four-dimensional snapshots (reflecting the microvoxel information and surface point cloud features of the cutting area at multiple time points) with the real-time acquired first tool parameters (including cutting speed, feed rate, depth of cut, tool model, and wear level) into a multidimensional input vector. This vector incorporates rich microstructural information that evolves in space and time, as well as the tool's current process and physical state. The causal hypernet then uses the node relationships and hypernet weights learned through causal graph constraints during the pre-training phase to perform forward reasoning on this input vector in the latent space, following the actual influence paths between the variables in the causal structure, and automatically predicts a set of optimal parameter adjustments. Finally, the server superimposes these adjustments with the original parameters to form a second set of tool parameters that is causally interpretable and physically reasonable, ready for real-time distribution to the machine tool.

[0057] The pre-configured deep learning causal supernet is a deep learning model architecture that combines a supernet mechanism with causal structure priors. It is specifically designed to generate parameter adjustment strategies for specific dynamic scenarios. It consists of two core components: a causal graph module, in which domain experts pre-construct a directed graph describing the causal relationships between key variables (such as cutting speed, feed rate, depth of cut, tool wear, material type, and processing temperature) based on machine tool processing technology and physical mechanisms, or through offline causal discovery algorithms; and a supernet module, a generative network that accepts a "target analysis vector" as input and outputs the main network parameters (or directly outputs parameter increments). This supernet is typically composed of a multi-layer fully connected network or graph neural network. Its weights and structure are pre-trained offline based on historical multi-condition data and causal graph constraints, also known as the "pre-configured" stage.

[0058] During the offline pre-setting process, the causal graph is first verified and fine-tuned using historical machining data from multiple batches, multiple materials, and multiple tool combinations to ensure that it reflects the core causal pathway from process input to machining quality output. Next, guided by this causal graph, the hypernetwork, through multi-task or meta-learning methods, learns a set of parameter sub-network weights that can quickly generate parameters that conform to physical and process logic. The training objectives include not only an optimization loss for final machining quality (such as surface roughness, vibration amplitude, and tool life), but also a causal consistency loss, forcing the hypernetwork-generated strategy to operate within the given causal structure.

[0059] During the online inference phase, after the system collects a new sequence of four-dimensional snapshots and current tool parameters, it normalizes and reduces their dimensionality, injects key causal node indicators, and then feeds them into a pre-trained deep learning causal supernet. The supernet uses its inherent causal constraints and empirical knowledge to quickly generate a set of adjustments, such as "reducing the cutting speed by 3m / min and increasing the feed rate by 0.02mm / rev." This output not only conforms to the causal influence path learned by the model in the preset phase, but also responds reasonably to the current dynamic working conditions. The entire process does not require retraining the main network, ensuring the efficiency and causal interpretability of online matching, thereby achieving accurate, reliable, and adaptive adjustment of tool parameters in complex and changing machining environments.

[0060] In one possible implementation, the four-dimensional snapshot sequence and the first tool parameters are input into a preset deep learning causal supernet to generate the second tool parameters, specifically including: normalizing and dimensionality reduction mapping the four-dimensional snapshot sequence and the first tool parameters to obtain a spatiotemporal feature summary and tool parameter features; concatenating the spatiotemporal feature summary and the tool parameter features into a unified vector, and appending key node indicators marked in a preset causal graph to obtain a target analysis vector, where the key node indicators include material type and ambient temperature; using a preset deep learning causal supernet constructed by a preset causal graph, combined with the supernet weights, to infer the target analysis vector to obtain an inference parameter adjustment amount; and superimposing the inference parameter adjustment amount and the first tool parameters to obtain the second tool parameters.

[0061] Specifically, the server first converts the previously obtained four-dimensional snapshot sequence and the first tool parameters into "feature summaries" that are easy for the model to process. Specifically, the voxel light intensity, phase distribution, and point cloud geometry information at each moment in the four-dimensional snapshot sequence are normalized to the same numerical range, and several key indicators representing the spatiotemporal evolution trend, such as the vibration amplitude change rate and micro-deformation rate, are extracted through dimensionality reduction mapping (such as principal component or autoencoder methods). At the same time, the cutting speed, feed rate, depth of cut, tool model, and wear level in the first tool parameters are also mapped into a set of ordered numerical features and similarly normalized. For example, if a gradual increase in vibration and the initial appearance of surface microcracks are observed during a section of aluminum alloy processing, the server will condense this information into a "spatiotemporal feature summary" and then map the original 150m / min cutting speed, 0.12mm / rev feed rate, 1.8mm depth of cut, and "TiAlN coated tool - moderate wear" into the corresponding parameter features.

[0062] Next, the server concatenates the two feature summaries into a complete vector and appends key node indicators such as "Material Type = Aluminum Alloy" and "Ambient Temperature = 22°C" to form the final target analysis vector. This vector is then fed into a deep learning causal supernet, trained offline based on a pre-set causal graph. The supernet's internal structure and weights incorporate causal path information, such as "processing speed → thermal deformation → vibration" and "wear → surface quality," enabling it to output an optimal set of parameter adjustments in a single forward inference. For example, based on current vibration trends and material properties, the model might recommend reducing the cutting speed by 10 m / min, increasing the feed rate by 0.02 mm / rev, and maintaining the same depth of cut. The server then adds these adjustments to the original parameters to generate new secondary tool parameters (i.e., 140 m / min, 0.14 mm / rev, 1.8 mm) and prepares to distribute them to the machine tool, enabling efficient, causal-driven adaptive matching of tool states.

[0063] In one possible implementation, refer to Figure 2 , Figure 2 Another schematic flow chart of a deep learning-based dynamic tool parameter matching method provided in an embodiment of the present application includes steps S210 to S220, which are as follows: S210: Receive custom tool parameters sent by a user for a machine tool cutting area; S220: Control the machine tool tool to dynamically match the custom tool parameters based on the custom tool parameters.

[0064] Specifically, when a user enters new parameters such as cutting speed, feed rate, or depth of cut on a monitoring interface or dedicated terminal, these custom instructions are packaged into standardized control messages and sent to the backend server. After receiving the message, the server first verifies the parameter range and safety (for example, it does not exceed the maximum allowable speed of the machine tool, the upper limit of the feed torque, etc.), and then converts the legal custom parameters into the corresponding CNC instruction format (usually by updating the server G-code server or issuing instant parameter adjustment commands through the server OPC-UA server interface), and pushes them to the server CNC server controller in real time. In this way, the system not only achieves seamless integration of user intent and machine tool control, but also performs necessary protection for parameter legitimacy before execution, ensuring that "human-machine collaboration" is both flexible and reliable.

[0065] For example, when machining aviation parts, an operator might observe minor surface scratches during a particular machining process. The operator then manually enters adjustment commands on the monitoring interface: "Reduce the cutting speed from 150 m / min to 135 m / min, and increase the feed rate from 0.12 mm / rev to 0.14 mm / rev." Upon receiving these commands, the server quickly verifies that these parameters are within the machine's tolerances and updates the corresponding G-code lines (e.g., G1 F0.14 S135) to the current machining program cache. The server then sends these updated commands to the machine's drive system via a real-time communication bus, and the tool immediately executes the cutting motion according to the new parameters at the start of the next cutting cycle. The entire process, from user input to machine response, often takes place within a few hundred milliseconds, ensuring rapid improvement in machining quality while maintaining precise control over the overall process.

[0066] This application also provides a tool parameter dynamic matching device based on deep learning, referring to Figure 3 , Figure 3 A module schematic diagram of a deep learning-based tool parameter dynamic matching device provided in an embodiment of the present application, wherein the device is a server, and the server includes an acquisition module 31 and a processing module 32, wherein the acquisition module 31 acquires laser holographic imaging data and three-dimensional contour scanning data for the machine tool cutting area at a target timestamp; the processing module 32 determines the holographic voxel data and cropped point cloud data based on the laser holographic imaging data and the three-dimensional contour scanning data; the processing module 32 inputs the holographic voxel data and the cropped point cloud data into a four-dimensional grid containing the target timestamp to obtain a four-dimensional snapshot sequence; the acquisition module 31 acquires the first tool parameter of the machine tool tool at the target timestamp; the processing module 32 inputs the four-dimensional snapshot sequence and the first tool parameter into a preset deep learning causal supernet to generate a second tool parameter.

[0067] In a possible implementation, the acquisition module 31 acquires the laser holographic imaging data and three-dimensional contour scanning data for the machine tool cutting area at the target timestamp, specifically including: the processing module 32 sends synchronization acquisition instructions to the laser holographic camera and the three-dimensional contour scanner respectively; the processing module 32 reads the timestamps returned by the laser holographic camera and the three-dimensional contour scanner, and determines that the timestamp is the target timestamp; under the constraint of the target timestamp, the processing module 32 receives the interference fringe sequence sent by the laser holographic camera and the point cloud batch sent by the three-dimensional contour scanner to obtain the laser holographic imaging data and the three-dimensional contour scanning data.

[0068] In a possible implementation, the processing module 32 determines the holographic voxel data and the cropped point cloud data based on the laser holographic imaging data and the three-dimensional contour scanning data, specifically including: the processing module 32 performs de-artifacting and brightness equalization on the interference fringe pattern contained in the laser holographic imaging data to obtain an initialized image; the processing module 32 maps the initialized image into three-dimensional voxels according to a predefined spatial grid, and each three-dimensional voxel includes a corresponding local phase and local light intensity; the processing module 32 performs statistical filtering on the contour scanning point cloud contained in the three-dimensional contour scanning data to obtain an initialized point cloud; the processing module 32 determines the surface normal on the initialized point cloud to obtain the intended area; the processing module 32 converts the three-dimensional voxel and the intended area into a unified coordinate and data structure respectively to obtain the holographic voxel data and the cropped point cloud data.

[0069] In one possible implementation, the processing module 32 inputs the holographic voxel data and the cropped point cloud data into a four-dimensional grid containing a target timestamp to obtain a four-dimensional snapshot sequence, specifically including: the acquisition module 31 obtains the engineering requirements of the machine tool cutting area; the processing module 32 sets the spatial three-dimensional resolution and the time step according to the engineering requirements; the processing module 32 fills the holographic voxel data into the four-dimensional grid according to the three-dimensional resolution to obtain a first moment slice; the processing module 32 fills the cropped point cloud data into the four-dimensional grid according to the time step to obtain a second moment slice; the processing module 32 splices the first moment slice and the second moment slice in chronological order to obtain a four-dimensional snapshot sequence.

[0070] In one possible implementation, the acquisition module 31 acquires the first tool parameters of the machine tool tool at the target timestamp, specifically including: the processing module 32 queries the cutting speed, feed rate and cutting depth corresponding to the machine tool tool at the target timestamp from the CNC controller; the processing module 32 synchronously acquires the tool model and wear level of the machine tool tool; the processing module 32 determines the first tool parameters based on the cutting speed, feed rate, cutting depth, tool model and wear level.

[0071] In one possible implementation, the processing module 32 inputs the four-dimensional snapshot sequence and the first tool parameters into a preset deep learning causal supernet to generate the second tool parameters, specifically including: the processing module 32 performs feature normalization and dimensionality reduction mapping on the four-dimensional snapshot sequence and the first tool parameters to obtain a spatiotemporal feature summary and tool parameter features; the processing module 32 splices the spatiotemporal feature summary and the tool parameter features into a unified vector, and appends the key node indicators marked in the preset causal graph to obtain a target analysis vector, where the key node indicators include material type and ambient temperature; the processing module 32 uses a preset deep learning causal supernet constructed by a preset causal graph, and combines the supernet weights to infer the target analysis vector to obtain an inference parameter adjustment amount; the processing module 32 superimposes the inference parameter adjustment amount and the first tool parameter to obtain the second tool parameter.

[0072] In a possible implementation, the acquisition module 31 receives the user-defined tool parameters sent for the machine tool cutting area; the processing module 32 controls the machine tool tool to dynamically match the customized tool parameters according to the customized tool parameters.

[0073] It should be noted that the above embodiments provide devices that implement their functions using only the division of the above functional modules as examples. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the device and method embodiments provided in the above embodiments are based on the same concept. The specific implementation process is detailed in the method embodiment and will not be repeated here.

[0074] This application also provides an electronic device, referring to Figure 4 , Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application. The electronic device may include: at least one processor 41, at least one network interface 44, a user interface 43, a memory 45, and at least one communication bus 42.

[0075] The communication bus 42 is used to realize the connection and communication between these components.

[0076] The user interface 43 may include a display screen (Display) and a camera (Camera). Optionally, the user interface 43 may also include a standard wired interface and a wireless interface.

[0077] The network interface 44 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).

[0078] The processor 41 may include one or more processing cores. Using various interfaces and circuits, the processor 41 connects to various components within the server. It executes instructions, programs, code sets, or instruction sets stored in the memory 45, as well as accesses data stored in the memory 45, to perform various server functions and process data. Optionally, the processor 41 may be implemented using at least one of the following hardware forms: a digital signal processing (DSP), a field-programmable gate array (FPGA), or a programmable logic array (PLA). The processor 41 may integrate one or a combination of a central processing unit (CPU), a graphics processing unit (GPU), and a modem. The CPU primarily processes the operating system, user interface, and application programs; the GPU is responsible for rendering and drawing content displayed on the display screen; and the modem handles wireless communications. It is understood that the modem may also be implemented as a separate chip, rather than integrated into the processor 41.

[0079] Among them, the memory 45 may include a random access memory (RAM) or a read-only memory (Read-Only Memory). Optionally, the memory 45 includes a non-transitory computer-readable storage medium. The memory 45 can be used to store instructions, programs, codes, code sets or instruction sets. The memory 45 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as a touch function, a sound playback function, an image playback function, etc.), instructions for implementing the above-mentioned various method embodiments, etc.; the data storage area may store data involved in the above-mentioned various method embodiments, etc. The memory 45 may also be optionally at least one storage device located away from the aforementioned processor 41. As Figure 4 As shown, the memory 45 as a computer storage medium may include an operating system, a network communication module, a user interface module, and an application program for a tool parameter dynamic matching method based on deep learning.

[0080] exist Figure 4In the electronic device shown, the user interface 43 is mainly used to provide an input interface for the user and obtain data input by the user; and the processor 41 can be used to call an application stored in the memory 45 that stores a tool parameter dynamic matching method based on deep learning. When executed by one or more processors, the electronic device executes one or more methods in the above embodiments.

[0081] It should be noted that for the aforementioned method embodiments, for simplicity of description, they are all expressed as a series of action combinations, but those skilled in the art should be aware that this application is not limited by the order of the actions described, because according to this application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required for this application.

[0082] The present application also provides a computer-readable storage medium storing instructions, which, when executed by one or more processors, enable an electronic device to execute one or more of the methods described in the above embodiments.

[0083] In the above embodiments, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0084] In the several embodiments provided in this application, it should be understood that the disclosed devices can be implemented in other ways. For example, the device embodiments described above are merely schematic, such as the division of units, which is only a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some service interface, and the indirect coupling or communication connection of devices or units can be electrical or other forms.

[0085] Units described as separate components may or may not be physically separate, and components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0086] In addition, the functional units in the various embodiments of the present application may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0087] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable memory. Based on this understanding, the technical solution of this application, or the portion that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the various embodiments of the method of this application. The aforementioned memory includes various media that can store program code, such as USB flash drives, mobile hard drives, magnetic disks, or optical disks.

[0088] The above is only an exemplary embodiment of the present disclosure and cannot be used to limit the scope of the present disclosure. That is, any equivalent changes and modifications made according to the teachings of the present disclosure are still within the scope of the present disclosure. After considering the disclosure of the specification and the truth of practice, those skilled in the art will easily think of other embodiments of the present disclosure. This application is intended to cover any variation, use or adaptive change of the present disclosure, which follows the general principles of the present disclosure and includes common knowledge or customary technical means in the art that are not recorded in the present disclosure. The description and examples are to be regarded as exemplary only, and the scope and spirit of the present disclosure are defined by the claims.

Claims

1. A tool parameter dynamic matching method based on deep learning, characterized in that: The method comprises: Acquire laser holographic imaging data and three-dimensional contour scanning data of the machine tool cutting area at the target time stamp; Determining holographic voxel data and cropped point cloud data according to the laser holographic imaging data and the three-dimensional contour scanning data; Inputting the holographic voxel data and the cropped point cloud data into a four-dimensional grid containing the target timestamp to obtain a four-dimensional snapshot sequence; Obtaining a first tool parameter of a machine tool at the target timestamp; The four-dimensional snapshot sequence and the first tool parameters are input into a preset deep learning causal supernet to generate second tool parameters.

2. The tool parameter dynamic matching method based on deep learning according to claim 1 is characterized in that: The obtaining of laser holographic imaging data and three-dimensional contour scanning data of the machine tool cutting area at a target time stamp specifically includes: Sending synchronous acquisition instructions to the laser holographic camera and the three-dimensional profile scanner respectively; Reading the timestamps returned by the laser holographic camera and the three-dimensional profile scanner, and determining that the timestamps are the target timestamps; Under the constraint of the target timestamp, the interference fringe sequence sent by the laser holographic camera and the point cloud batch sent by the three-dimensional contour scanner are received to obtain the laser holographic imaging data and the three-dimensional contour scanning data.

3. The tool parameter dynamic matching method based on deep learning according to claim 1, characterized in that: The determining of holographic voxel data and cropped point cloud data based on the laser holographic imaging data and the three-dimensional contour scanning data specifically includes: Performing artifact removal and brightness equalization on the interference fringe pattern contained in the laser holographic imaging data to obtain an initialized image; Mapping the initialization image into three-dimensional voxels according to a predefined spatial grid, each of the three-dimensional voxels including a corresponding local phase and local intensity; Performing statistical filtering on the contour scanning point cloud contained in the three-dimensional contour scanning data to obtain an initialized point cloud; Determining a surface normal on the initialization point cloud to obtain an intended region; The three-dimensional voxel and the intended area are converted into unified coordinates and data structures respectively to obtain the holographic voxel data and the cropped point cloud data.

4. The tool parameter dynamic matching method based on deep learning according to claim 1, characterized in that: Inputting the holographic voxel data and the cropped point cloud data into a four-dimensional grid containing the target timestamp to obtain a four-dimensional snapshot sequence specifically includes: Obtaining engineering requirements for the cutting area of ​​the machine tool; Set the spatial 3D resolution and time step according to the engineering requirements; Filling the holographic voxel data into the four-dimensional grid according to the three-dimensional resolution to obtain a first time slice; Filling the cropped point cloud data into the four-dimensional grid according to the time step to obtain a second time slice; The first moment slice and the second moment slice are sequentially spliced ​​in chronological order to obtain the four-dimensional snapshot sequence.

5. The tool parameter dynamic matching method based on deep learning according to claim 1, characterized in that: The obtaining of the first tool parameter of the machine tool at the target timestamp specifically includes: Querying the cutting speed, feed rate, and cutting depth of the machine tool corresponding to the target timestamp from the CNC controller; Simultaneously obtaining the tool model and wear grade of the machine tool tool; The first tool parameters are determined according to the cutting speed, feed rate, cutting depth, tool type and wear level.

6. The tool parameter dynamic matching method based on deep learning according to claim 1, characterized in that: Inputting the four-dimensional snapshot sequence and the first tool parameters into a preset deep learning causal supernet to generate second tool parameters specifically includes: performing feature normalization and dimensionality reduction mapping on the four-dimensional snapshot sequence and the first tool parameter to obtain a spatiotemporal feature summary and tool parameter features; splicing the spatiotemporal feature summary and the tool parameter feature into a unified vector, and appending key node indicators marked in a preset causal graph to obtain a target analysis vector, wherein the key node indicators include material type and ambient temperature; Using the preset deep learning causal supernet constructed by the preset causal graph, and combining the supernet weights to reason on the target analysis vector, to obtain an inference parameter adjustment amount; The inference parameter adjustment amount and the first tool parameter are superimposed to obtain the second tool parameter.

7. The tool parameter dynamic matching method based on deep learning according to claim 1, characterized in that: The method further comprises: receiving user-defined tool parameters sent by the user for the cutting area of ​​the machine tool; According to the user-defined tool parameters, the machine tool tool is controlled to dynamically match the user-defined tool parameters.

8. A tool parameter dynamic matching device based on deep learning, characterized in that: The device comprises an acquisition module (31) and a processing module (32), wherein: The acquisition module (31) is used to acquire laser holographic imaging data and three-dimensional contour scanning data of the machine tool cutting area at a target time stamp; The processing module (32) is used to determine holographic voxel data and cropped point cloud data based on the laser holographic imaging data and the three-dimensional contour scanning data; The processing module (32) is further configured to input the holographic voxel data and the cropped point cloud data into a four-dimensional grid containing the target timestamp to obtain a four-dimensional snapshot sequence; The acquisition module (31) is further used to acquire a first tool parameter of a machine tool at the target time stamp; The processing module (32) is further configured to input the four-dimensional snapshot sequence and the first tool parameters into a preset deep learning causal supernet to generate second tool parameters.

9. An electronic device, characterized in that: The electronic device comprises a processor (41), a memory (45), a user interface (43) and a network interface (44), wherein the memory (45) is used to store instructions, the user interface (43) and the network interface (44) are both used to communicate with other devices, and the processor (41) is used to execute the instructions stored in the memory (45) so that the electronic device executes the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores instructions, and when the instructions are executed, the method according to any one of claims 1 to 7 is performed.

Images