A horizontal CNC machine tool and its control method

By acquiring and storing historical data of the entire lifecycle of horizontal CNC machine tools, quantifying multi-dimensional scene features, generating standard feature fingerprints, and utilizing scene intelligent recognition modules and wear prediction models, precise adaptation of tool wear compensation and fault early warning are achieved, solving the problem of insufficient tool wear compensation accuracy in existing technologies.

CN123113673APending Publication Date: 2026-11-17TONGXIANG FOSHAN PRECISION MACHINERY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611420289.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-14
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing CNC machine tool wear compensation methods cannot adapt to different business scenarios such as roughing/finishing, different workpiece materials, and batch/single-piece customization. The compensation accuracy fluctuates greatly, making it difficult to achieve accurate matching for specific scenarios.

Method used

By acquiring and storing historical data of the entire lifecycle of a horizontal CNC machine tool, quantifying multi-dimensional scene features, generating standard feature fingerprints for the scene library, matching real-time features using a scene intelligent recognition module, constructing a multi-dimensional wear prediction model, outputting tool compensation values ​​and feeding them back to the controller, and achieving precise tool wear compensation.

Benefits of technology

It improves the accuracy of tool wear compensation, achieves precise adaptation to different business scenarios, reduces scrap rate, and enables early warning and adaptive monitoring of tool wear failure.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123113673A_ABST
    Figure CN123113673A_ABST
Patent Text Reader

Abstract

This invention relates to the field of CNC machine tool technology, and provides a horizontal CNC machine tool and its control method. The method includes: quantifying multi-dimensional scene features contained in historical data to generate standard feature fingerprints for scene libraries of different dimensions; matching real-time features of the CNC machine tool with the standard feature fingerprints to obtain real-time business scenarios; training and fitting a constructed wear prediction model using historical batch tool wear and product size deviation data under the same scenario as a training set, outputting tool compensation values, and generating compensation instructions based on the tool compensation values. This invention acquires and stores historical data throughout the entire lifecycle of a horizontal CNC machine tool, quantifies the multi-dimensional scene features contained in the historical data, and generates standard feature fingerprints for scene libraries of different dimensions. Based on these standard feature fingerprints, it can achieve intelligent identification of different dimensional business scenarios, thereby achieving accurate adaptation of wear compensation and improving the accuracy of tool wear compensation.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of CNC machine tool technology, and more specifically, to a horizontal CNC machine tool and its control method. Background Technology

[0002] Horizontal CNC machine tools are mainly composed of a bed, column, worktable, spindle system, feed transmission system and CNC system. They are particularly suitable for machining parts such as boxes and shells, and can complete multi-face machining in one clamping.

[0003] For CNC machine tools, tool wear will occur after prolonged use. The presence of wear will affect the accuracy and quality of workpiece machining. For CNC machine tool wear compensation, the usual compensation methods generally use a fixed wear rate model or uniform compensation parameters. These methods cannot adapt to different business scenarios such as roughing / finishing, different workpiece materials, batch / single-piece customization, etc. The compensation accuracy fluctuates greatly and it is difficult to achieve accurate matching for specific scenarios. Summary of the Invention

[0004] Based on this, in order to improve the accuracy of tool wear compensation in CNC machine tools, the present invention provides a horizontal CNC machine tool and its control method, the specific technical solution of which is as follows: A horizontal CNC machine tool, comprising: The CNC machine tool data center is equipped with a historical database, which is used to acquire and store historical data of the entire life cycle of horizontal CNC machine tools, and quantify the multi-dimensional scene features contained in the historical data to generate standard feature fingerprints of scene libraries in different dimensions. The scene intelligent recognition module is used to acquire real-time features of CNC machine tools, match real-time features with standard feature fingerprints, and obtain real-time business scenes; The multi-dimensional wear prediction model corresponds one-to-one with multiple business scenarios. It is used to train and fit the historical batch tool wear and product size deviation data under the same scenario as the training set, output tool compensation value and feed it back to the controller. The controller receives tool compensation values ​​and generates compensation commands that can be executed by the CNC system based on these values.

[0005] The horizontal CNC machine tool acquires and stores historical data throughout its entire lifecycle, quantifies the multi-dimensional scene features contained in the historical data, and generates standard feature fingerprints for different dimensions of scene libraries. Based on these standard feature fingerprints, it can achieve intelligent recognition of different business scenarios, thereby achieving precise adaptation of wear compensation and improving the accuracy of tool wear compensation.

[0006] Preferably, the scene intelligent recognition module includes: The similarity acquisition unit is used to acquire the similarity between real-time features and standard feature fingerprints, and to determine whether there is at least one similarity greater than a preset similarity threshold. If there is, the business scenario corresponding to the highest similarity greater than the similarity threshold is taken as the real-time business scenario; otherwise, a new scenario entry is created. The new scenario migration unit is used to obtain the feature similarity between the new scenario and the existing business scenario, and to generate the initial wear prediction model of the new scenario based on the wear prediction models of the several existing business scenarios with the highest feature similarity.

[0007] Preferably, the horizontal CNC machine tool further includes: The wear fault early warning model is used to obtain the real-time wear rate of CNC machine tool tools and the historical normal wear rate of tools under the same business scenario. Based on the real-time wear rate and the historical normal wear rate, the relative wear deviation is obtained, and the tool wear fault early warning monitoring is performed based on the relative wear deviation.

[0008] Preferably, the CNC machine tool data center divides scene features into four categories, and uses differentiated coding methods to quantify different categories of scene features to generate standard feature fingerprints for scene libraries of different dimensions.

[0009] Preferably, the different categories of scene features include hard classification features, ordered level features, core numerical features, and weakly influential numerical features. One-hot encoding is used to quantize hard classification features, ordered level features are used to quantize ordered level features, logarithmic normalization is used to quantize core numerical features, and linear normalization is used to quantize weakly influential numerical features.

[0010] Preferably, the similarity acquisition unit uses the maximum mutual information coefficient to quantify the correlation between each feature and the wear rate, calculates the dynamic weight of the correlation similarity after normalization, and performs candidate scene matching and screening based on hard classification features. For candidate scenes, the similarity between real-time features and standard feature fingerprints is calculated based on dynamic weights.

[0011] A control method for a horizontal CNC machine tool, applied to the horizontal CNC machine tool, includes the following steps: S1: Acquire and store historical data of the entire life cycle of the horizontal CNC machine tool, and quantify the multi-dimensional scene features contained in the historical data to generate standard feature fingerprints of scene libraries in different dimensions. S2, acquire the real-time features of the CNC machine tool, match the real-time features with standard feature fingerprints to obtain the real-time business scenario; S3. Construct a multi-dimensional wear prediction model. Use historical batch tool wear and product size deviation data under the same scenario as the training set to train and fit the constructed wear prediction model, output tool compensation value, and generate compensation instructions that can be executed by the CNC system based on the tool compensation value.

[0012] Preferably, the horizontal CNC machine tool control method further includes the following steps: S4, obtain the similarity between real-time features and standard feature fingerprints, determine whether there is at least one similarity greater than a preset similarity threshold, and if not, create a new scene entry; S5: Obtain the feature similarity between the new scenario and the existing business scenarios, and generate the initial wear prediction model for the new scenario based on the wear prediction models of the several existing business scenarios with the highest feature similarity.

[0013] Preferably, step S1 specifically includes: S11 divides scene features into four categories: hard classification features, ordered level features, core numerical features, and weakly influential numerical features. S12, different coding methods are used to quantify the features of different categories of scenes to generate standard feature fingerprints of scene libraries in different dimensions; Specifically, one-hot encoding is used to quantize hard classification features, ordinal encoding is used to quantize ordered level features, logarithmic normalization is used to quantize core numerical features, and linear normalization is used to quantize weakly influential numerical features.

[0014] Preferably, step S3 specifically includes: S31. Obtain core data of historical batches under the same scenario, perform benchmark alignment on the core data to form data samples, perform process consistency screening, abnormal sample removal and multi-batch sample fusion on the data samples to obtain the average wear sequence and size deviation sequence under the same scenario, and construct a wear history data training set based on the average wear sequence and size deviation sequence. S32 divides the tool's entire life cycle wear into the initial wear stage, the normal wear stage, and the rapid wear stage. For different wear stages, a wear function model with corresponding physical mechanism is adopted, and the first boundary point between the initial wear stage and the normal wear stage and the second boundary point between the normal wear stage and the rapid wear stage are solved based on the curvature characteristics of the wear rate. S33 fits the wear function model for different wear stages using a training set of wear history data, and decouples the size direction and compensation type corresponding to wear according to the process type. S34. For different wear stages, a mapping model between wear amount and size deviation is constructed, and the mapping model is fitted by the wear history data training set. Based on the fitted wear function model and the mapping model, the actual size deviation is obtained, and the tool compensation value is obtained by reverse calculation based on the actual size deviation. The core data include cumulative cutting mileage, wear amount, and dimensional deviation. Attached Figure Description

[0015] The invention will be further understood from the following description taken in conjunction with the accompanying drawings. The components in the drawings are not necessarily drawn to scale, but rather the emphasis is on illustrating the principles of the embodiments. In different views, the same reference numerals designate corresponding parts.

[0016] Figure 1 This is a schematic diagram of the overall process of a control method for a horizontal CNC machine tool according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating step S1 in one embodiment of the present invention; Figure 3 This is a flowchart illustrating step S3 in one embodiment of the present invention; Figure 4 This is a schematic diagram of the overall process of a control method for a horizontal CNC machine tool according to another embodiment of the present invention; Figure 5 This is a schematic diagram of an overall system with a horizontal CNC machine tool according to one embodiment of the present invention; Figure 6 This is a schematic diagram of the functional unit structure of the scene intelligent recognition module in one embodiment of the present invention; Figure 7 This is a schematic diagram of the overall structure of a horizontal CNC machine tool according to an embodiment of the present invention.

[0017] Explanation of reference numerals in the attached drawings: 1. Bed; 2. Spindle box; 3. Feed drive; 4. Control panel. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to its embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not limit the scope of protection of the invention.

[0019] It should be noted that when an element is referred to as being "fixed to" another element, it can be directly attached to the other element or there may be an intervening element. When an element is referred to as being "connected to" another element, it can be directly connected to the other element or there may be an intervening element. The terms "vertical," "horizontal," "left," "right," and similar expressions used herein are for illustrative purposes only and do not represent the only possible implementation.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to limit the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0021] In this invention, "first" and "second" do not represent a specific quantity or order, but are merely used to distinguish names.

[0022] like Figure 5 As shown, one embodiment of the present invention provides a horizontal CNC machine tool, including a CNC machine tool data center, a scene intelligent recognition module, a multi-dimensional wear prediction model, and a controller. In addition, as... Figure 7 As shown, a horizontal CNC machine tool generally includes several major parts such as the bed (1), spindle box (2), feed transmission device (3), operating table (4), tool post, tailstock, CNC system, cooling system, and chip removal system. The bed, as the basic support component of the machine tool, is usually made of cast iron and undergoes aging treatment. Internally, it uses rationally distributed ribs to ensure rigidity and stability. Horizontal lathe beds are divided into flat beds and slant beds; slant beds are more conducive to chip removal and maintaining accuracy.

[0023] The spindle box contains the spindle, speed change mechanism, and bearings, responsible for driving the workpiece rotation to complete the main motion. The workpiece is clamped at the front end of the spindle via a chuck or fixture interface, with common connection methods including synchronous belt drive and gearbox drive. The feed drive is responsible for the longitudinal (Z-axis) and transverse (X-axis) movement of the tool post. CNC machine tools typically use servo motors to directly drive ball screws, resulting in a short transmission chain and high precision. This is also the component that differs most significantly from the structure of a conventional horizontal lathe.

[0024] The tool post is used to install and change tools. CNC lathes are mostly equipped with electric or hydraulic automatic indexing tool posts, enabling multiple machining operations to be completed in a single setup. The tailstock is used to support the tail end of long workpieces and can also be used to install tools such as centers and drills. It plays a crucial role when machining long shaft parts. On ordinary horizontal lathes, the lead screw is dedicated to thread machining, while the feed screw handles general cutting feed; on CNC machine tools, ball screws are mostly used instead.

[0025] The CNC machine tool data center has a historical database, which is used to acquire and store historical data of the entire life cycle of horizontal CNC machine tools, and quantify the multi-dimensional scene features contained in the historical data to generate standard feature fingerprints of scene libraries in different dimensions.

[0026] CNC machine tool data centers include, but are not limited to, servers and cloud computing. Historical databases can be used to store data such as tool wear records, machine tool accuracy calibration records, fault case libraries, process parameter libraries, and product quality libraries. Data cleaning, alignment, and standardization are performed for easy retrieval and use later. Internal data of the CNC system typically includes spindle power, feed position, and torque, while lathe data includes temperature and vibration. Since the structure of horizontal CNC machine tools and the acquisition and storage of their operating parameters are standard technical aspects, they will not be elaborated upon here.

[0027] Multi-dimensional scenario features include four dimensions: process, production, equipment, and environment. Process-related scenario features include, but are not limited to, tool type / material, workpiece material, the three cutting elements (rotation speed, feed rate, depth of cut), and operation type (milling / boring / drilling / grinding). Production-related scenario features include production batch size, accuracy level, machining cycle time, and tool change cycle. Equipment-related scenario features include equipment age, cumulative runtime, current load range, and recent maintenance records. Environmental-related scenario features include ambient temperature and humidity.

[0028] The scene intelligent recognition module is used to acquire real-time features of CNC machine tools, match real-time features with standard feature fingerprints, and obtain real-time business scenes.

[0029] As a preferred technical solution, such as Figure 6 As shown, the scene intelligent recognition module includes a similarity acquisition unit and a new scene migration unit.

[0030] The similarity acquisition unit is used to obtain the similarity between real-time features and standard feature fingerprints, and to determine whether there is at least one similarity greater than a preset similarity threshold, such as 85%. If so, the business scenario corresponding to the highest similarity greater than the similarity threshold is used as the real-time business scenario; otherwise, a new scenario entry is created.

[0031] Business scenarios include, but are not limited to, heavy-duty roughing, high-precision finishing, intermittent cutting, long-cycle batch production, single-piece trial production, and equipment aging processing. For different business scenarios, corresponding dedicated fault feature libraries are established to match the sensitive characteristics of different fault types. For example, for roughing scenarios, key monitoring can be conducted on sudden changes in cutting force and vibration peaks, corresponding to chipping and tool breakage faults; for finishing scenarios, key monitoring can be conducted on high-frequency components of acoustic emission and dimensional deviation fluctuations, corresponding to tool micro-chipping and early spindle bearing failures; in intermittent cutting scenarios, key monitoring can be conducted on torque fluctuations and feed impacts, corresponding to tool loosening and leadscrew wear faults.

[0032] The new scenario migration unit is used to obtain the feature similarity between the new scenario and the existing business scenario, and to generate the initial wear prediction model for the new scenario based on the wear prediction models of the several existing business scenarios with the highest feature similarity.

[0033] For example, wear prediction models from 3-5 existing business scenarios with the highest feature similarity can be selected, and the model parameters can be weighted and fused according to their corresponding similarity to obtain an initial wear prediction model for the new scenario. In this way, as processed data accumulates, rapid iterative optimization through incremental learning can solve the cold start problem of models in small sample scenarios.

[0034] The multi-dimensional wear prediction model corresponds one-to-one with multiple business scenarios. It is used to train and fit the model using historical batch tool wear and product size deviation data under the same scenario as a training set, output tool compensation values ​​and feed them back to the controller. The controller is used to receive the tool compensation values ​​and generate compensation instructions that can be executed by the CNC system based on the tool compensation values.

[0035] Specifically, tool wear curves from historical batches under the same scenario can be extracted—that is, the curve function showing the change in flank wear as a function of cutting mileage / time. A standard wear baseline is established using piecewise nonlinear fitting for the three stages of tool wear: initial wear, normal wear, and rapid wear. Simultaneously, by associating product dimensional deviation data under the corresponding scenario, a mapping model between wear and dimensional deviation is established. Finally, by combining the standard wear baseline and the mapped wear, and obtaining the current cutting mileage / time, the corresponding tool compensation value is obtained, and tool length compensation and radius compensation commands that can be executed by the CNC system are directly output.

[0036] In summary, the horizontal CNC machine tool acquires and stores historical data throughout its entire lifecycle, quantifies the multi-dimensional scene features contained in the historical data, and generates standard feature fingerprints for different dimensions of scene libraries. Based on these standard feature fingerprints, it can achieve intelligent recognition of different business scenarios, thereby achieving precise adaptation of wear compensation and improving the accuracy of tool wear compensation.

[0037] In one embodiment, the horizontal CNC machine tool further includes a wear fault early warning model.

[0038] The wear fault early warning model is used to obtain the real-time wear rate of CNC machine tool tools and the historical normal wear rate of tools under the same business scenario. Based on the real-time wear rate and the historical normal wear rate, the relative wear deviation is obtained, and the tool wear fault early warning monitoring is performed based on the relative wear deviation.

[0039] Different fault warning monitoring intervals are set, and tool wear fault warning monitoring is performed based on the monitoring interval and the relative wear deviation. For example, when the relative wear deviation is between 0.8 and 1.2, it can be defined as the normal interval, in which the tool wear is in the abnormal interval of the normal fluctuation range; when the relative wear deviation is between 1.2 and 1.5, it is defined as the abnormal interval, in which the tool wear is accelerated and should be alerted; when the relative wear deviation is between 1.5 and 2.0, it is defined as the warning interval, in which the tool is at risk of chipping, cracking, or transmission failure; the fault interval: when the relative wear deviation is > 2.0, it is defined as the fault interval, in which the tool wears rapidly or has already failed. It is adapted to different business scenarios and can realize adaptive warning monitoring of tool wear faults.

[0040] In this example, by obtaining the relative wear deviation, tool wear is combined with fault early warning, which makes the fault judgment more consistent with the actual state of the equipment, and at the same time realizes early monitoring and early warning of faults, reducing the scrap rate.

[0041] In one embodiment, the CNC machine tool data center divides scene features into four categories, quantifies different categories of scene features using differentiated coding methods, and generates standard feature fingerprints for scene libraries of different dimensions.

[0042] Specifically, different categories of scene features include hard classification features, ordered level features, core numerical features, and weakly influential numerical features. One-hot encoding is used to quantize hard classification features, ordered level features are used to quantize ordered level features, logarithmic normalization is used to quantize core numerical features, and linear normalization is used to quantize weakly influential numerical features.

[0043] Hard classification features include, but are not limited to, process type (milling / boring / drilling), tool category, and workpiece material category. They employ one-hot encoding and hard-matching masks, with a perfect match resulting in a 1 and otherwise a 0, serving as a coarse screening condition. Ordered grade features include, but are not limited to, accuracy grade (IT6~IT12), production batch grade, and equipment service life grade. These can be encoded using grade ordinal encoding, with similarity calculated based on grade differences, decreasing by 20% for each grade difference. Core numerical features include, but are not limited to, the three cutting elements (speed, feed, depth of cut), cutting force, and hardness. These are processed using logarithmic normalization and cosine similarity calculation, resulting in the highest matching accuracy. Weakly influential numerical features include, but are not limited to, ambient temperature, coolant temperature, and spindle load rate. These can be processed using linear normalization and cosine similarity calculation, with lower weights.

[0044] Since cutting parameters and wear rate have an exponential relationship (referencing Taylor's tool life formula), linear normalization would amplify the distance deviation in the high parameter range. Therefore, logarithmic normalization, rather than nonlinear normalization, is used for the core numerical features. After mapping to logarithmic space, the relative changes in parameters correspond linearly to the relative changes in wear rate, allowing similarity calculations to better reflect the physical laws of wear.

[0045] As a preferred technical solution, the similarity acquisition unit uses the maximum mutual information coefficient to quantify the correlation between each feature and the wear rate, calculates the dynamic weight of the correlation similarity after normalization, and performs candidate scene matching and screening based on hard classification features. For candidate scenes, the similarity between real-time features and standard feature fingerprints is calculated based on dynamic weights to obtain the comprehensive similarity.

[0046] Since different features have vastly different degrees of influence on wear characteristics, equal weighting would cause the matching results to deviate from the core contradiction. Therefore, the maximum mutual information coefficient is used here to quantify the correlation between each feature and the wear rate, and after normalization, it is used as the dynamic weight for similarity calculation.

[0047] Specifically, based on the full historical data, the MIC (Maximal Information Coefficient) value of each feature and the steady-state wear rate in the corresponding scenario is calculated to characterize the contribution of the feature to wear. Then, the MIC values ​​of all features are normalized to the range of [0,1] to obtain the weight coefficients of each dimension, and the sum of the weight coefficients of each dimension is 1. Finally, the weights of each dimension can be dynamically updated according to the major category of the scenario. For example, the dimension feature weight of the depth of cut is the highest in the roughing scenario, and the dimension feature weight of the feed rate is the highest in the finishing scenario.

[0048] First, masking is performed using hard classification features, and fine-grained matching is conducted only within the scene subset of the same process and tool category to compress the number of candidate scenes and adapt to the computing performance of the CNC system's edge. Based on this, the weighted cosine similarity between the numerical features of the candidate scenes and the real-time features is calculated. Finally, the weighted cosine similarity and the similarity between the real-time features and the ordered hierarchical features are weighted and fused to obtain the final comprehensive similarity.

[0049] like Figure 1 As shown, an embodiment of the present invention also provides a control method for a horizontal CNC machine tool, applied to the horizontal CNC machine tool, comprising the following steps: S1 acquires and stores historical data of the entire lifecycle of the horizontal CNC machine tool, and quantifies the multi-dimensional scene features contained in the historical data to generate standard feature fingerprints of scene libraries in different dimensions.

[0050] As a preferred technical solution, such as Figure 2As shown, step S1 specifically includes: S11 divides scene features into four categories: hard classification features, ordered level features, core numerical features, and weakly influential numerical features.

[0051] S12 quantizes different categories of scene features using differentiated encoding methods to generate standard feature fingerprints for scene libraries of different dimensions.

[0052] Specifically, one-hot encoding is used to quantize hard classification features, ordinal encoding is used to quantize ordered level features, logarithmic normalization is used to quantize core numerical features, and linear normalization is used to quantize weakly influential numerical features.

[0053] S2: Obtain the real-time features of the CNC machine tool, match the real-time features with standard feature fingerprints, and obtain the real-time business scenario.

[0054] It can obtain the similarity between real-time features and standard feature fingerprints, and determine whether there is at least one similarity greater than a preset similarity threshold, such as 85%. If so, the business scenario corresponding to the highest similarity greater than the similarity threshold is taken as the real-time business scenario.

[0055] S3. Construct a multi-dimensional wear prediction model. Use historical batch tool wear and product size deviation data under the same scenario as the training set to train and fit the constructed wear prediction model, output tool compensation value, and generate compensation instructions that can be executed by the CNC system based on the tool compensation value.

[0056] As a preferred technical solution, such as Figure 3 As shown, step S3 specifically includes: S31. Obtain core data of historical batches under the same scenario, perform benchmark alignment on the core data to form data samples, perform process consistency screening, abnormal sample removal and multi-batch sample fusion on the data samples to obtain the average wear sequence and size deviation sequence under the same scenario, and construct a wear history data training set based on the average wear sequence and size deviation sequence.

[0057] The data samples are screened for process consistency, and batches with cutting parameter fluctuations exceeding ±5%, abnormal coolant, or interrupted tool changes are removed to ensure that all samples are in the same process steady state. Then, abnormal samples are removed, which can be done using the 3σ criterion. Samples with wear deviations from the mean by more than 3 times the standard deviation at the same mileage, such as chipped edges, rolled edges, and inspection errors, are removed to obtain qualified samples. Finally, multiple batches of samples are fused. All qualified samples are grouped at equal intervals according to cutting mileage. The median of wear and dimensional deviation is taken for each group to obtain the average wear sequence and dimensional deviation sequence under the same scenario, eliminating the interference of individual tool discreteness.

[0058] S32 divides the tool's entire life cycle wear into the initial wear stage, the normal wear stage, and the rapid wear stage. For different wear stages, a wear function model with corresponding physical mechanism is adopted, and the first boundary point between the initial wear stage and the normal wear stage and the second boundary point between the normal wear stage and the rapid wear stage are solved based on the curvature characteristics of the wear rate.

[0059] Based on the tribology theory of metal cutting, the wear of a cutting tool throughout its entire life cycle is divided into three stages, each employing a nonlinear model corresponding to its physical mechanism. Specifically, in the initial wear stage, the microscopic protrusions on the cutting edge of a new tool are rapidly smoothed, the contact area increases exponentially, and the wear rate decreases. Therefore, an exponential decay model can be used. First, the initial microscopic wear amount of the cutting edge and the total wear increment during the running-in period are obtained. Then, the exponential decay wear rate is obtained based on the ratio of the cumulative cutting mileage to a preset running-in mileage constant. Finally, the wear amount is obtained based on the wear rate, the initial microscopic wear amount of the cutting edge, and the total wear increment during the running-in period. For example, the wear amount = initial microscopic wear amount of the cutting edge + total wear increment during the running-in period × (1 - exp(-real-time cumulative cutting mileage / running-in mileage constant)).

[0060] During the normal wear stage, abrasive wear and adhesive wear alternate steadily, with the wear rate being approximately constant. It slowly accelerates as the cutting edge becomes dull. At this time, the wear rate can be represented by a linear model, for example, as: Wear amount = Wear amount at the end of the running-in period + Steady-state wear rate × (Real-time cumulative cutting mileage - First dividing point) + Slow acceleration coefficient × (Real-time cumulative cutting mileage - First dividing point) 2 The steady-state wear rate and the slow acceleration coefficient can both be set empirically or based on experimental calibration.

[0061] During the rapid wear stage, the cutting edge becomes severely dull, cutting force and temperature increase dramatically, thermal wear dominates, and the wear rate rises exponentially. At this point, the wear rate can be modeled using an exponential growth model. For example, wear amount = critical dulling value for normal wear × exp(rapid wear acceleration coefficient × (real-time cumulative cutting mileage - second dividing point)). The rapid wear acceleration coefficient can be set empirically or based on experimental calibration.

[0062] Local weighted regression smoothing can be applied to the fused average wear sequence to eliminate detection noise, resulting in a continuous smooth curve. The first derivative (wear rate) and second derivative (wear acceleration / curvature) can then be calculated. Generally, the initial wear rate of a tool exhibits a decreasing characteristic, and the normal wear rate tends to stabilize, approximately equal to zero. Therefore, the first boundary point corresponds to the minimum wear rate, i.e., the first zero-crossing point where the second derivative changes from negative to positive. Note that the first boundary point must fall within the 10%-30% range of the total mileage; if it does not, the smoothing window should be adaptively adjusted and recalculated. In the later stages of normal wear, the tool wear rate begins to accelerate. The second boundary point corresponds to the maximum wear acceleration, and the wear rate corresponding to the second boundary point must be greater than a certain value, such as 1.5 times the average tool wear rate in the normal stage. The second boundary point must fall within the 70%-90% range of the total mileage to ensure compliance with the general law of tool wear.

[0063] To avoid compensation jumps caused by abrupt changes in wear rate at segmentation points, a dual constraint of continuous function value and continuous first derivative can be applied during fitting. The constrained Levenberg-Marquardt algorithm is then used to solve the model parameters, thereby obtaining a globally smooth and continuous standard wear baseline to characterize the individual discreteness of the tool.

[0064] S33 fits the wear function model for different wear stages using a training set of historical wear data, and decouples the size direction and compensation type corresponding to wear according to the process type.

[0065] As shown in the table below: Process type Direction of wear Affecting workpiece dimensions Corresponding CNC compensation type Theoretical transmission relationship External turning Radial unilateral wear workpiece diameter Tool radius compensation Diameter deviation ≈ 2 × single-sided wear amount Internal boring Radial unilateral wear Workpiece hole diameter Tool radius compensation Hole diameter deviation ≈ -2 × single-sided wear amount End face milling (depth) Axial wear Processing depth Tool length compensation Depth deviation ≈ Axial wear Contour milling (side view) Radial wear Outline dimensions Tool radius compensation Lateral deviation ≈ Radial wear S34 constructs a mapping model between wear amount and dimensional deviation for different wear stages, and fits the mapping model with a training set of wear history data. Based on the fitted wear function model and the mapping model, the actual dimensional deviation is obtained, and the tool compensation value is obtained by reverse calculation based on the actual dimensional deviation. The core data include cumulative cutting mileage, wear amount and dimensional deviation.

[0066] The cumulative cutting mileage, used as the abscissa of wear, can be calculated by integrating the spindle speed, feed rate, and cutting time recorded in real time by the CNC system, thus eliminating the interference of speed fluctuations on the wear rate. For wear amount, the flank wear detection data for the entire lifecycle of each batch of tools can be extracted, including on-machine visual inspection and offline microscopic inspection data. Each wear detection value is aligned to the corresponding cumulative cutting mileage node, forming discrete wear sample points. Simultaneously, the key dimension detection values ​​of the machined workpiece at the corresponding wear node are extracted, and the dimensional deviation is calculated as actual dimension - theoretical dimension, which is then paired one-to-one with the wear sample points.

[0067] After obtaining the dimensional deviation from the mapping model, the compensation increment, i.e., the tool compensation value, is calculated in reverse according to the CNC system's compensation logic. The tool compensation value can be directly written into the CNC system's tool offset table. It can be understood that the compensation goal is to eliminate the current dimensional deviation and bring the machined dimension back to its theoretical value. For axial dimensions such as milling depth and turning end face length, length compensation can be used; for radial dimensions such as outer diameter, inner hole, and side profile, radius compensation is used.

[0068] Thus, by selecting a segmented model based on the physical mechanism of wear and automatically identifying the boundary point through curvature extrema, and performing global fitting with continuity constraints, the accuracy of dimensional deviation prediction in the wear stage can be improved.

[0069] In summary, the horizontal CNC machine tool control method acquires and stores historical data throughout the entire lifecycle of the horizontal CNC machine tool, quantifies the multi-dimensional scene features contained in the historical data, and generates standard feature fingerprints for different dimensions of scene libraries. Based on the standard feature fingerprints of the scene library, it can achieve intelligent recognition of different business scenarios, thereby achieving accurate adaptation of wear compensation and improving the accuracy of tool wear compensation.

[0070] In one embodiment, such as Figure 4 As shown, the horizontal CNC machine tool control method further includes the following steps: S4. Obtain the similarity between real-time features and standard feature fingerprints, and determine whether there is at least one similarity greater than a preset similarity threshold. If not, create a new scene entry.

[0071] S5: Obtain the feature similarity between the new scenario and the existing business scenarios, and generate the initial wear prediction model for the new scenario based on the wear prediction models of the several existing business scenarios with the highest feature similarity.

[0072] An improved Arcard wear model can be used as a unified physical framework, decomposing the model parameters into two categories: general physical parameters and scenario-sensitive parameters. To address the transferability of different parameters, corresponding business scenarios are retrieved from different similarity dimensions for parameter transfer. Finally, calibration is completed through Bayesian updates using a very small number of samples, generating a complete and usable initial wear prediction model.

[0073] Specifically, the improved Archard wear model, widely validated in engineering, is adopted as the basic framework to anchor the fundamental physical laws of wear. This improved Archard wear model outputs the improved Archard wear model, with cutting speed, feed rate, depth of cut, and cutting stroke as variables. Model parameters include the comprehensive wear coefficient and the wear index. The specific function form of the improved Archard wear model can be appropriately set based on the actual scenario, without excessive limitations here. The wear index, as a general physical parameter, is determined by the wear mechanism of the material pair, exhibits stable values ​​within the same material system, and has strong transferability, generally showing small fluctuations when the scenario changes. The comprehensive wear absorption is a scenario-sensitive parameter; the comprehensive wear coefficient, also a scenario-sensitive parameter, is affected by multiple factors such as cutting parameters, machining method, cooling conditions, and equipment status, resulting in significant scenario differences and serving as the core object for transfer correction.

[0074] For general physical parameters, the feature similarity between the new scenario and existing business scenarios can be calculated and weighted for the two characteristics of tool material and workpiece material, resulting in a first matching degree. Then, the wear indices corresponding to several wear prediction models with the highest first matching degree, such as three existing business scenario wear prediction models, are selected, and the arithmetic mean of each type of wear index is used as the initial wear index of the initial wear prediction model for the new scenario. For scenario-sensitive parameters, the comprehensive similarity of cutting parameters, cooling method, and process type can be obtained, and the comprehensive wear coefficients corresponding to several wear prediction models with the highest comprehensive similarity, such as five existing business scenario wear prediction models, are selected. The comprehensive wear coefficients are then weighted and averaged based on the comprehensive similarity and used as the initial wear coefficient of the initial wear prediction model for the new scenario.

[0075] For the initial wear prediction model obtained for the new scenario, perform small-sample Bayesian incremental calibration and update.

[0076] In this way, the feature similarity between the new scenario and the existing business scenario is obtained, and the initial wear prediction model of the new scenario is generated based on the wear prediction models of the existing business scenarios with the highest feature similarity. This not only makes use of the value of historical data, but also avoids the physical contradictions caused by pure data migration, and reduces the sample size required for cold start.

[0077] Generally, spindle power and vibration signals simultaneously contain no-load components, operating condition fluctuation components, and actual wear components. Directly using the measured spindle power and vibration signals for wear estimation can misjudge operating condition interferences such as cutting parameter fluctuations and uneven material hardness as wear changes, increasing the misjudgment rate of dynamic compensation. Therefore, an offline calibrated process reference model can be used to separate the no-load and operating condition fluctuation components from the real-time signal, extracting pure wear-sensitive features and eliminating non-wear interferences at the source.

[0078] Specifically, in the same scenario, observable signals such as spindle power and vibration can be decomposed into three independent components: the equipment no-load component, the process condition component, and the wear-related component. The equipment no-load component (such as spindle idle power and machine tool foundation vibration) is completely unrelated to the cutting process and tool wear; the process condition component is determined by the three cutting elements and workpiece material properties, and only reflects the magnitude of the cutting load, which is unrelated to tool wear; the wear-related component is generated by the increase in cutting force and frictional vibration caused by tool edge dulling, and is a direct and linear representation of the wear rate.

[0079] Based on this, a benchmark mapping model of process condition components and cutting parameters is fitted offline using historical steady-state machining data from the same scenario, serving as the benchmark for feedforward decoupling. For the power benchmark model, a power function form consistent with the Taylor tool life formula can be adopted to ensure physical consistency. It uses cutting speed, feed rate, and depth of cut as variable parameters and power as the output. For the vibration benchmark model, the vibration acceleration RMS value can be selected as the basic feature, and the 1-5kHz mid-to-high frequency vibration energy can be extracted simultaneously as auxiliary features to establish the corresponding vibration benchmark model.

[0080] Before each batch of machining, when the spindle is running unloaded until its speed stabilizes, the spindle's unloaded power and vibration are collected. The theoretical power and vibration components are obtained based on the power and vibration reference models, respectively. Then, based on the real-time total spindle power, unloaded spindle power, and theoretical power components, the pure wear power increment (i.e., real-time total spindle power - unloaded spindle power - theoretical power component) is obtained. Based on the real-time total vibration acceleration RMS value, unloaded vibration, and theoretical vibration components, the pure wear vibration increment (i.e., real-time total vibration acceleration RMS value - unloaded vibration - theoretical vibration component) is obtained. The high-frequency energy increment during wear is also calculated (i.e., real-time total high-frequency energy - unloaded high-frequency reference energy - theoretical high-frequency energy). Finally, based on the current cumulative cutting mileage L, the two boundary points of the standard wear baseline are matched to determine the current wear stage in real time, and the corresponding stage's feature weights and stage correction gains are automatically applied.

[0081] The high-frequency reference energy under no-load conditions can be obtained by collecting and calculating the average energy of the same frequency band when the spindle is running stably before machining. The total high-frequency energy during machining can be obtained by performing a sliding window FFT on the real-time time-domain vibration signal to obtain the spectrum, extracting the spectral lines in the target frequency band, and calculating the sum of squared amplitudes (energy integral). The high-frequency energy in the process theory can be obtained by offline calibration of the reference model of cutting parameters and high-frequency energy under the same scenario, and then substituting it into the real-time parameters for calculation.

[0082] Feature weights can be preset. The wear mechanisms differ at different wear stages, and the correlation between each feature and the wear rate varies significantly. Based on historical data from the same scenario, the maximum mutual information coefficient between the feature and the wear rate at each stage is calculated, normalized, and used as the adaptive feature weight.

[0083] For example, in the initial wear stage, the feature weights of the pure wear power increment, the pure wear vibration increment, and the high-frequency energy increment during wear are set to 0.35, 0.45, and 0.2, respectively; in the normal wear stage, they can be set to 0.6, 0.25, and 0.15, respectively; and in the rapid wear stage, they can be set to 0.3, 0.3, and 0.4, respectively. Based on the feature weights, the increments are normalized and weighted to obtain the comprehensive wear deviation, which characterizes the degree of wear of the current tool relative to the standard baseline.

[0084] A stage correction gain is introduced (0.6 for the initial wear stage, 0.9 for the normal wear stage, and 0.7 for the rapid wear stage) to suppress excessive correction in unstable stages, resulting in a real-time wear rate correction factor: Real-time wear rate correction factor = 1 + Overall wear deviation × Stage correction gain. If the real-time wear rate correction factor > 1, it indicates that the current tool wear is faster than the standard baseline, and the compensation amount needs to be increased; if the real-time wear rate correction factor = 1, it indicates that the previous tool wear is completely consistent with the standard baseline, and compensation is based on the original baseline value; if the real-time wear rate correction factor < 1, it indicates that the current tool wear is slower than the standard baseline, and the compensation amount can be appropriately reduced.

[0085] As a preferred technical solution, the real-time wear rate can be obtained by correcting the theoretical wear rate of the standard wear baseline, and the current actual wear amount can be approximated using the sliding window trapezoidal integral method: Current actual wear amount = Standard wear baseline value at the last calibration point + Corrected real-time wear rate + Cumulative wear increment. Sampling is performed at fixed mileage intervals to obtain the correction factor and standard wear rate at the i-th sampling point. The product of the fixed mileage interval, correction factor, and standard wear rate is used as the wear increment within a single sampling interval. Finally, the cumulative wear increment is the sum of the wear increments across all sampling intervals. The total sampling mileage is the cumulative cutting mileage from the last calibration mileage to the current cumulative cutting mileage.

[0086] In this way, by decomposing the real-time machining signal into three independent components—no-load, working condition, and wear—and offsetting non-wear interference through feedforward of the process reference model, the accuracy of wear rate estimation can be improved. Based on the differences in wear mechanisms, feature weight adaptive matching can be performed in stages, which can improve the scenario adaptability of tool wear compensation.

[0087] The technical features of the embodiments described can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0088] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A horizontal CNC machine tool, characterized in that, include: The CNC machine tool data center is equipped with a historical database, which is used to acquire and store historical data of the entire life cycle of horizontal CNC machine tools, and quantify the multi-dimensional scene features contained in the historical data to generate standard feature fingerprints of scene libraries in different dimensions. The scene intelligent recognition module is used to acquire real-time features of CNC machine tools, match real-time features with standard feature fingerprints, and obtain real-time business scenes; The multi-dimensional wear prediction model corresponds one-to-one with multiple business scenarios. It is used to train and fit the historical batch tool wear and product size deviation data under the same scenario as the training set, output tool compensation value and feed it back to the controller. The controller receives tool compensation values ​​and generates compensation commands that can be executed by the CNC system based on these values.

2. A horizontal CNC machine tool as described in claim 1, characterized in that, The scene intelligent recognition module includes: The similarity acquisition unit is used to acquire the similarity between real-time features and standard feature fingerprints, and to determine whether there is at least one similarity greater than a preset similarity threshold. If there is, the business scenario corresponding to the highest similarity greater than the similarity threshold is taken as the real-time business scenario; otherwise, a new scenario entry is created. The new scenario migration unit is used to obtain the feature similarity between the new scenario and the existing business scenario, and to generate the initial wear prediction model of the new scenario based on the wear prediction models of the several existing business scenarios with the highest feature similarity.

3. A horizontal CNC machine tool as described in claim 2, characterized in that, Also includes: The wear fault early warning model is used to obtain the real-time wear rate of CNC machine tool tools and the historical normal wear rate of tools under the same business scenario. Based on the real-time wear rate and the historical normal wear rate, the relative wear deviation is obtained, and the tool wear fault early warning monitoring is performed based on the relative wear deviation.

4. A horizontal CNC machine tool as described in claim 3, characterized in that, The CNC machine tool data center divides scene features into four categories and uses differentiated coding methods to quantify the scene features of different categories, generating standard feature fingerprints for scene libraries of different dimensions.

5. A horizontal CNC machine tool as described in claim 4, characterized in that, Different categories of scene features include hard classification features, ordered level features, core numerical features, and weakly influential numerical features. One-hot encoding is used to quantize hard classification features, ordered level features are used to quantize ordered level features, logarithmic normalization is used to quantize core numerical features, and linear normalization is used to quantize weakly influential numerical features.

6. A horizontal CNC machine tool as described in claim 5, characterized in that, The similarity acquisition unit uses the maximum mutual information coefficient to quantify the correlation between each feature and the wear rate, calculates the dynamic weights based on the normalized correlation similarity, and performs candidate scene matching and screening based on hard classification features. For candidate scenes, the similarity between real-time features and standard feature fingerprints is calculated based on the dynamic weights.

7. A control method for a horizontal CNC machine tool, applied to the horizontal CNC machine tool as described in any one of claims 1-6, characterized in that, Includes the following steps: S1: Acquire and store historical data of the entire life cycle of the horizontal CNC machine tool, and quantify the multi-dimensional scene features contained in the historical data to generate standard feature fingerprints of scene libraries in different dimensions. S2, acquire the real-time features of the CNC machine tool, match the real-time features with standard feature fingerprints to obtain the real-time business scenario; S3. Construct a multi-dimensional wear prediction model. Use historical batch tool wear and product size deviation data under the same scenario as the training set to train and fit the constructed wear prediction model, output tool compensation value, and generate compensation instructions that can be executed by the CNC system based on the tool compensation value.

8. The horizontal CNC machine tool control method as described in claim 7, characterized in that, It also includes the following steps: S4, obtain the similarity between real-time features and standard feature fingerprints, determine whether there is at least one similarity greater than a preset similarity threshold, and if not, create a new scene entry; S5: Obtain the feature similarity between the new scenario and the existing business scenarios, and generate the initial wear prediction model for the new scenario based on the wear prediction models of the several existing business scenarios with the highest feature similarity.

9. The horizontal CNC machine tool control method as described in claim 8, characterized in that, Step S1 specifically includes: S11 divides scene features into four categories: hard classification features, ordered level features, core numerical features, and weakly influential numerical features. S12, different coding methods are used to quantify the features of different categories of scenes to generate standard feature fingerprints of scene libraries in different dimensions; Specifically, one-hot encoding is used to quantize hard classification features, ordinal encoding is used to quantize ordered level features, logarithmic normalization is used to quantize core numerical features, and linear normalization is used to quantize weakly influential numerical features.

10. The horizontal CNC machine tool control method as described in claim 9, characterized in that, Step S3 specifically includes: S31. Obtain core data of historical batches under the same scenario, perform benchmark alignment on the core data to form data samples, perform process consistency screening, abnormal sample removal and multi-batch sample fusion on the data samples to obtain the average wear sequence and size deviation sequence under the same scenario, and construct a wear history data training set based on the average wear sequence and size deviation sequence. S32 divides the tool's entire life cycle wear into the initial wear stage, the normal wear stage, and the rapid wear stage. For different wear stages, a wear function model with corresponding physical mechanism is adopted, and the first boundary point between the initial wear stage and the normal wear stage and the second boundary point between the normal wear stage and the rapid wear stage are solved based on the curvature characteristics of the wear rate. S33 fits the wear function model for different wear stages using a training set of wear history data, and decouples the size direction and compensation type corresponding to wear according to the process type. S34. For different wear stages, a mapping model between wear amount and size deviation is constructed, and the mapping model is fitted by the wear history data training set. Based on the fitted wear function model and the mapping model, the actual size deviation is obtained, and the tool compensation value is obtained by reverse calculation based on the actual size deviation. The core data include cumulative cutting mileage, wear amount, and dimensional deviation.