Numerical control machining process identification method and system based on multi-axis state signal

Abstract

The application discloses a kind of based on multi-axis state signal's numerical control processing technology identification method and system, belong to machining process type identification technical field.Method includes to workpiece is standardized clamping positioning;Make numerical control machine tool execute standardization preoperation program;Adopt non-synchronous acquisition mode based on working condition trigger, obtain multi-axis current data;Introduce timestamp alignment algorithm to realize the consistency of multi-axis current data on time sequence;For the multi-axis current data of timestamp alignment, combined with empirical mode decomposition and wavelet threshold denoising algorithm, construct adaptive selective denoising preprocessing framework, determine the current data of sample segmentation, and utilize multi-axis synchronous sliding window technology, obtain three-dimensional data set;Process feature set is constructed;Adopt principal component analysis to process feature set is nonlinear feature dimension reduction, and combined with grid optimization's radial basis function kernel support vector mechanism constructs multi-class process identification model.The application effectively realizes the automatic identification of multiple processing technology.

Description

Technical Field

[0001] The present invention relates to a method and system for identifying a numerical control machining process based on multi-axis state signals, and belongs to the technical field of identifying mechanical machining process types. Background Art

[0002] In modern manufacturing, the accurate identification and real-time monitoring of machining processes are of great significance for improving production efficiency, optimizing the operating state of equipment, and reducing production costs. Especially in multi-axis controlled numerical control machining equipment, different machining processes such as profile milling, arc machining, inclined line machining, etc. have a direct impact on machining quality and equipment performance. However, traditional methods mainly stay at the identification of machining conditions and mainly rely on manual experience or simple threshold judgment, which are difficult to meet the requirements of multi-axis linkage in complex machining scenarios. For example, an operator judges the condition state of a certain axis through the experience of machining sound, vibration or current fluctuation, or monitors the single-axis signal parameters through a preset threshold. Although these methods are simple to implement, they have obvious deficiencies in terms of consistency, accuracy and adaptability. Especially in multi-variety and small-batch production, manual judgment is inefficient, and the threshold method is difficult to effectively distinguish the subtle changes between processes.

[0003] At present, in order to improve the accuracy of process identification, some technical solutions attempt to obtain machining state information by collecting internal data of the machine tool such as spindle speed, feed rate, tool position, etc., or installing various external sensors such as vibration sensors, force sensors, acoustic emission sensors, etc. Although collecting internal data of the machine tool can provide rich operating parameters, it usually needs to be deeply integrated into the numerical control system, involving complex interface development and data synchronization problems, and has poor compatibility with different models of machine tools. In addition, the installation of external sensors not only increases the hardware cost and maintenance difficulty, but also makes the data processing process cumbersome due to the variety of signal types, restricting its wide application in actual production. When analyzing the characteristics of machining processes, these methods are often limited to the signal characteristics of a single dimension, and do not fully utilize the comprehensive information of multi-channel data, and it is difficult to systematically reveal the differences between different processes and the commonalities of the same process.

[0004] In view of this, the present invention is specifically proposed. Summary of the Invention

[0005] This invention provides a method and system for CNC machining process identification based on multi-axis state signals. By installing easily deployable current sensors on the machining equipment and using asynchronous acquisition methods to collect current data under multi-axis control to construct a process feature set, the data acquisition process can be significantly simplified, avoiding direct dependence on machine tool internal data and the integration requirements of complex sensor systems. Furthermore, principal component analysis is used to perform nonlinear feature dimensionality reduction on the feature set, and a multi-class process identification model is constructed by combining a support vector machine with grid-optimized radial basis function kernels, thereby realizing the automatic classification and identification of various machining processes.

[0006] The technical solution of this invention is:

[0007] According to a first aspect of the present invention, a method for identifying CNC machining processes based on multi-axis state signals is provided, comprising:

[0008] S1. Based on the CNC machine tool processing technology type, the current sensors are respectively placed on the U-phase lines of the servo drive devices of the spindle and the required feed axis in the electrical cabinet of the CNC machine tool, and each current sensor is connected to the data acquisition card for data acquisition.

[0009] S2. Before each formal machining operation of a workpiece on the CNC machine tool, the workpiece is clamped and positioned in a standardized manner; then the CNC machine tool is executed a standardized pre-run program.

[0010] S3. The spindle with the current sensor installed and the required feed axis are regarded as the first type of axis. For the first type of axis, under different types of machining processes, the asynchronous acquisition method based on working condition triggering is adopted to obtain multi-axis current data respectively. A timestamp alignment algorithm is introduced to achieve the consistency of multi-axis current data in time sequence and obtain multi-axis current data after timestamp alignment.

[0011] S4. An adaptive selective noise reduction preprocessing framework is constructed by combining empirical mode decomposition and wavelet threshold noise reduction algorithm to make decisions on the current data of each axis in the timestamp-aligned multi-axis current data and the reconstructed current data, and the current data of each axis determined by the decision is used as the current data to be segmented as samples.

[0012] S5. Based on the current data to be segmented, a three-dimensional dataset is obtained using multi-axis synchronous sliding window technology;

[0013] S6. Extract multi-dimensional time-frequency domain features from the sliding window samples of the three-dimensional dataset under different types of processing technology, and construct a process feature set;

[0014] S7. Principal component analysis is used to perform nonlinear feature dimensionality reduction on the process feature set, and a multi-class process identification model is constructed by combining a support vector machine with a grid-optimized radial basis function kernel.

[0015] Furthermore, S2 specifically includes:

[0016] S21. Before each formal machining of a workpiece on a CNC machine tool, the workpiece shall be clamped and positioned in a standardized manner using standard fixtures.

[0017] S22. Enable the CNC machine tool to execute a standardized pre-run program:

[0018] Spindle idling and preheating stage: The spindle is run under no-load self-test to allow the spindle motor to reach a stable operating temperature and eliminate electrical system abnormalities.

[0019] Low-speed no-load reciprocating motion stage of each feed axis: Under no-load conditions, perform a low-speed reciprocating motion self-check on each feed axis;

[0020] Medium-speed no-load reciprocating motion stage of each feed axis: Under no-load conditions, each feed axis performs medium-speed reciprocating motion, so that the corresponding servo motor completes the medium-speed reciprocating motion self-check and achieves parameter stability.

[0021] Return to machine origin operation phase: Reset each feed to the reference initial position;

[0022] Self-inspection stage at the beginning of machining: By executing a unified and standardized operation process for spindle lifting and lowering, the consistency of the state at the beginning of machining is ensured.

[0023] Furthermore, the standardized workpiece clamping and positioning specifically includes: selecting a clamp that ensures the contact area between the workpiece and the clamp reaches more than 60% based on the workpiece material characteristics and geometry; tightening the clamp at a preset position using a predetermined clamping force to ensure that the clamping error detected by a coordinate measuring machine or indicator meets the preset standard.

[0024] Furthermore, the method of obtaining multi-axis current data by adopting an asynchronous acquisition method based on working condition triggering is as follows: the end of the unified and standardized operation process of the spindle lifting and lowering the tool is used as the trigger condition for starting the current sensor; after the trigger condition is met, the current sensors of the spindle and the required feed axis are started to acquire the current signal with the same number of sampling points, thereby obtaining multi-axis current data.

[0025] Furthermore, S4 specifically includes:

[0026] Empirical mode decomposition is performed on the current data of each axis in the timestamp-aligned multi-axis current data to obtain multiple IMF components;

[0027] If any IMF component satisfies any judgment criterion based on frequency domain energy distribution characteristics or statistical properties, it is identified as a noise IMF component.

[0028] The wavelet threshold denoising algorithm is applied only to the identified noisy IMF components, while keeping the other IMF components unchanged, to obtain the reconstructed current data.

[0029] The energy retention rate of the reconstructed current data is examined, and a decision is made based on the energy retention rate: if the energy retention rate is lower than a preset threshold, the original current data before reconstruction is determined to be the current data to be segmented; otherwise, the reconstructed current data is determined to be the current data to be segmented.

[0030] Furthermore, S5 specifically includes:

[0031] S51. Based on the current data to be segmented from the sample, establish a multi-axis data matrix. The expression is:

[0032]

[0033] in, It refers to the number of shafts in the first type of shaft. This is the total number of sampling points for each axis. Indicating the first type of axis The first axis corresponds to the current data to be segmented by the sample. One sampling point; The matrix dimension is ;

[0034] S52, Set the length of the sliding window and sliding step size The sliding window is slid according to the set length. For each slide, the current data of all axes in the first type of axis within the same sliding window are extracted to form a sliding window sample. The expression is:

[0035]

[0036] in, Indicates the starting point index of the window is Sliding window samples, The index of the starting point of the window;

[0037] S53, Based on sliding step length By continuously sliding the window to a set length, multiple sliding window samples are generated from the original multi-axis data matrix, resulting in a three-dimensional dataset. , represented as:

[0038]

[0039] in, Indicates co-generation There are 1 sliding window sample, and the size of each sliding window sample is 1. .

[0040] Furthermore, the determination of the sliding window length and sliding step size is as follows: first, define one cycle using the current data of the spindle sample to be divided; calculate the number of sampling points A corresponding to one cycle under the minimum spindle speed process parameters, and take the smallest power of 2 greater than A as the sliding window length; take the sliding step size as more than half of the sliding window length.

[0041] Furthermore, S7 specifically includes:

[0042] S71. Standardize each feature in the process feature set to obtain the standardized feature matrix;

[0043] S72. Principal component analysis is used on the standardized features, with the goal of explaining more than 95% of the cumulative variance. Singular value decomposition is used to calculate the principal components and determine the number of principal components. The standardized feature matrix is ​​then projected onto the principal component space to generate a dimensionality-reduced feature matrix.

[0044] S73. Based on the dimensionality reduction feature matrix, random sampling is performed according to a preset ratio to divide it into a training set and a test set, while maintaining a balanced proportion of each processing technology category.

[0045] S74. Initialize the support vector machine model, select the radial basis function kernel RBF, set the regularization parameter and kernel coefficient, and enable the probability estimation function.

[0046] S75. Use the training set to fit the support vector machine model, optimize the hyperplane to maximize the inter-class margin, and obtain a multi-class process recognition model; use the test set to evaluate the performance of the multi-class process recognition model and verify its classification ability.

[0047] According to a second aspect of the present invention, a CNC machining process identification system based on multi-axis state signals is provided, comprising a module of the CNC machining process identification method based on multi-axis state signals as described in any one of the preceding claims.

[0048] The beneficial effects of this invention are:

[0049] I. Before machining a workpiece, this invention performs standardized clamping and positioning, and executes a standardized pre-running program that includes an unloaded running phase. This effectively eliminates the transient characteristics of CNC machine tools, ensuring that each axis is in a stable and controllable initial state. This process significantly reduces data deviations caused by mechanical system instability, guaranteeing the reliability and consistency of subsequent data acquisition from the source, and laying a solid foundation for accurate process identification.

[0050] Second, this invention employs a low-cost, asynchronous acquisition method suitable for low-frequency current signals. It can effectively acquire multi-axis current data without a unified hardware clock, avoiding the complex integration requirements of traditional methods for machine tool internal data such as spindle speed and feed rate. Furthermore, it eliminates the need for expensive and cumbersome external sensors such as vibration and force sensors. This design significantly reduces hardware complexity and implementation costs, while improving compatibility with different models of CNC machining equipment, making process identification technology easier to promote and apply in small and medium-sized manufacturing enterprises.

[0051] III. This invention proposes a selective noise reduction preprocessing workflow based on a combination of Empirical Mode Decomposition (EMD) and wavelet thresholding. This is used to improve the usability and stability of current signals in CNC machining processes. The method first uses EMD to decompose the original current signal into multiple Intrinsic Mode Components (IMFs). Then, frequency domain energy distribution analysis automatically identifies IMF components that may contain high-frequency noise, and wavelet thresholding is applied to these components for targeted suppression of non-stationary noise. Compared to traditional fixed-parameter methods such as static bandpass filtering, this method has stronger signal adaptability, maintaining high signal fidelity and feature consistency under different machining conditions, providing a more reliable signal foundation for subsequent machining state identification and feature extraction.

[0052] IV. This invention employs multi-axis synchronous sliding window technology to extract time-frequency domain features from multi-axis current data, including RMS value, peak value, variance, skewness, kurtosis, peak factor, dominant frequency amplitude, and spectral entropy, to construct a comprehensive process feature set. Principal component analysis is combined to achieve nonlinear dimensionality reduction of the feature set, and a multi-class process identification model is constructed using a grid-optimized radial basis function kernel support vector machine. By fully utilizing the collaborative information of multi-axis signals, the identification accuracy for typical processes such as X-shaped milling, Y-shaped milling, circular arc machining, and oblique line machining is significantly improved, demonstrating excellent robustness and adaptability.

[0053] In summary, this invention addresses the problems of poor data consistency, weak feature robustness, low classification accuracy, and complex system deployment in existing technologies by simplifying the current data acquisition process, employing a selective noise reduction method based on EMD and wavelet thresholding, a feature extraction mechanism covering various statistical and frequency domain information, and combining PCA dimensionality reduction and SVM classification models into a comprehensive recognition framework. This method not only improves the recognition accuracy and system stability of various typical processing techniques but also offers advantages such as low cost, no need for external sensors, and applicability to multiple equipment models. Furthermore, by accurately identifying the stable state of the processing process, this invention can provide high-quality data support for equipment operation monitoring, process optimization, and health management, demonstrating good engineering adaptability and industrial promotion value. Attached Figure Description

[0054] Figure 1 This is a flowchart of the method of the present invention;

[0055] Figure 2 This is a system architecture diagram of the present invention;

[0056] Figure 3 Example 2 is a flowchart of an adaptive selective noise reduction preprocessing algorithm that combines empirical mode decomposition (EMD) with wavelet threshold noise reduction.

[0057] Figure 4 This is the prediction result diagram for Example 2. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be arbitrarily combined with each other.

[0059] Example 1: As Figures 1-2 As shown, according to a first aspect of the present invention, a method for identifying CNC machining processes based on multi-axis state signals is provided, comprising:

[0060] S1. Based on the CNC machine tool processing technology type, place the current sensors on the U-phase lines of the servo drive devices of the spindle and the required feed axis in the electrical cabinet of the CNC machine tool, and connect each current sensor to the data acquisition card.

[0061] S2. Before each formal machining of a workpiece on the CNC machine tool, the workpiece is clamped and positioned in a standardized manner; then the CNC machine tool is executed with a standardized pre-run program to eliminate the transient characteristics of the CNC machine tool and ensure that each axis of the CNC machine tool is in a controllable initial state.

[0062] CNC machine tools are mainly composed of electrical systems, mechanical systems, and other systems.

[0063] The mechanical system involves each feed axis and the spindle. For the spindle: during the idle preheating phase, it is first run unloaded, driven by the motor, to allow the motor to reach a stable operating temperature and perform a self-check of the electrical system; during the machining start-up self-check phase, the standardized "tool lift-down" action is completed to confirm the consistent initial tool position. For each feed axis: in the low-speed phase: under no-load conditions, each feed axis performs low-speed reciprocating motion to detect and eliminate backlash in mechanical transmission components (such as lead screw-nut pairs, couplings, and guide rails), confirming that the mechanical structure is free from looseness or jamming; in the medium-speed phase: under no-load conditions, each feed axis performs medium-speed reciprocating motion, allowing the servo motor and driver to complete dynamic characteristic self-checks, stabilizing the current loop, speed loop, and position loop parameters, and verifying that the servo system has no alarms or oscillations.

[0064] The electrical system involves servo drive devices and sensors. The servo drive devices are responsible for sending drive signals to the spindle motor and each feed axis servo motor to realize actions such as preheating, self-testing, and homing. The sensors detect the spindle temperature, position, speed, and the origin and stroke signals of the feed axes in real time, providing closed-loop feedback to the driver to ensure the accuracy of preheating, backlash elimination, and origin reset.

[0065] Furthermore, S2 specifically includes:

[0066] S21. Before each formal machining of a workpiece on a CNC machine tool, use standard fixtures to perform standardized clamping and positioning of the workpiece to ensure clamping rigidity and positioning accuracy, and effectively eliminate the influence of clamping stress on process type identification.

[0067] S22. Enable the CNC machine tool to execute a standardized pre-run program:

[0068] Spindle idling and preheating stage: The spindle is run under no-load self-test to allow the spindle motor to reach a stable operating temperature and eliminate electrical system abnormalities.

[0069] Low-speed no-load reciprocating motion stage of each feed axis: Under no-load conditions, perform a low-speed reciprocating motion self-check on each feed axis to eliminate the backlash of each feed axis;

[0070] Medium-speed no-load reciprocating motion stage of each feed axis: Under no-load conditions, each feed axis performs medium-speed reciprocating motion, so that the corresponding servo motor completes the medium-speed reciprocating motion self-check and achieves parameter stability.

[0071] Return to machine origin operation phase: Reset each feed to the reference initial position;

[0072] Self-inspection stage at the beginning of machining: By executing a unified and standardized operation process for spindle lifting and lowering, the consistency of the state at the beginning of machining is ensured.

[0073] The standardized workpiece clamping and positioning specifically includes: selecting a clamp that ensures the contact area between the workpiece and the clamp reaches more than 60% based on the workpiece's material properties and geometry; tightening the clamp at a preset position using a predetermined clamping force; and ensuring that the clamping error detected by a coordinate measuring machine or indicator meets the preset standard.

[0074] S3. The spindle with the current sensor installed and the required feed axis are regarded as the first type of axis. For the first type of axis, under different types of machining processes, the asynchronous acquisition method based on working condition triggering is adopted to obtain multi-axis current data. A timestamp alignment algorithm is introduced to achieve the consistency of multi-axis current data in time sequence and obtain multi-axis current data after timestamp alignment. Taking CNC milling machine as an example, the machining process types include: X-shaped milling, Y-shaped milling, circular arc milling, and oblique line milling.

[0075] For the first type of axis, an asynchronous acquisition method based on working condition triggering is adopted to obtain multi-axis current data. Specifically, the end of the unified and standardized operation process of the spindle lifting and lowering the tool is used as the trigger condition for starting the current sensor. After the trigger condition is met, the current sensors of the spindle and the required feed axis are started to acquire the current signal with the same number of sampling points to obtain multi-axis current data.

[0076] As can be seen from the above, the present invention cleverly applies an asynchronous acquisition method based on working condition triggering to realize the acquisition of current signals of the main spindle and feed axis respectively. When using asynchronous acquisition of current signals of each axis of CNC machine tool, multiple channels share one ADC of the acquisition card, eliminating the need for an independent ADC for each axis, thus reducing costs. The servo motor current is in the low frequency range below kHz, making the asynchronous time difference negligible, and eliminating the need to pay extra costs for strict phase synchronization.

[0077] Furthermore, the timestamp alignment algorithm in S3 includes:

[0078] S31. Select the primary axis as the reference axis from the multi-axis current data, and the remaining axes as the second type of axes; use the current data of the reference axis as the reference signal;

[0079] S32. Add a timestamp with millisecond precision to each sampling point of the multi-axis current data to ensure that the sampling timing is traceable;

[0080] Assign a timestamp to each sampling point The calculation formula is:

[0081]

[0082] in, The initial sampling time, Indicates the first The timestamp of each sampling point For sampling point index, Sampling frequency (unit: Hz).

[0083] S33. Calculate the system delay of each axis in the second type of axis relative to the reference signal through correlation analysis. :

[0084]

[0085] in, The first reference signal One sampling point, For the second type of axis The first axis current data One sampling point, This is the time offset.

[0086] Subsequently, based on the calculated delay For the The timestamps of the current data for each axis are compensated and adjusted.

[0087]

[0088] in, For the first The first axis The original timestamps of each sampling point For the first The first axis The corrected timestamp of each sampling point after delay compensation.

[0089] It should be noted that this method is applicable when the current signals of each axis exhibit coordinated motion or similar dynamic response characteristics within a corresponding time period. For axes that do not show significant current changes or are not involved in motion control during this period, the alignment operation can be skipped to avoid introducing invalid or erroneous hysteresis compensation due to invalid correlation calculations. (For example, during X-axis linear milling, if the Y-axis is stationary, its current signal may not participate in the alignment process at this stage.)

[0090] S4. An adaptive selective noise reduction preprocessing framework is constructed by combining empirical mode decomposition (EMD) and wavelet threshold denoising algorithm to make decisions on the current data of each axis in the timestamp-aligned multi-axis current data and the reconstructed current data, and the current data of each axis determined by the decision is used as the current data to be segmented as samples.

[0091] Furthermore, S4 specifically includes:

[0092] Empirical Mode Decomposition (EMD) is performed on the current data of each axis in the timestamp-aligned multi-axis current data to obtain multiple IMF components;

[0093] If any IMF component satisfies any judgment criterion based on frequency domain energy distribution characteristics or statistical properties, it is identified as a noise IMF component.

[0094] The wavelet threshold denoising algorithm is applied only to the identified noisy IMF components, while keeping the other IMF components unchanged, to obtain the reconstructed current data.

[0095] The energy retention rate of the reconstructed current data is examined, and a decision is made based on the energy retention rate: if the energy retention rate is lower than a preset threshold, the original current data before reconstruction is determined to be the current data to be segmented; otherwise, the reconstructed current data is determined to be the current data to be segmented.

[0096] First, EMD decomposition is performed on the current signals of each axis, adaptively decomposing these nonlinear and non-stationary signals into a sequence of intrinsic mode functions (IMFs) with clear physical meaning. Then, by analyzing the frequency domain energy distribution characteristics and statistical properties of each IMF, high-frequency components containing noise interference are intelligently identified. Next, a noise discrimination mechanism determines which IMFs are indeed noise before applying wavelet thresholding for fine noise reduction, selectively suppressing random high-frequency interference while fully preserving key process feature components. Subsequently, the processed IMFs are reconstructed into a complete signal, and an energy loss verification mechanism checks the energy difference between the reconstructed result and the original signal. If the difference is too large, the reconstruction is abandoned to avoid filtering out useful features by simply reducing noise. Thus, while preserving the true process information to the greatest extent, noise interference is effectively eliminated, providing current input data with high signal-to-noise ratio, complete features, and clear physical meaning for subsequent process identification tasks.

[0097] S5. Based on the current data to be segmented, a three-dimensional dataset is obtained using multi-axis synchronous sliding window technology;

[0098] Furthermore, S5 specifically includes:

[0099] S51. Based on the current data to be segmented from the sample, establish a multi-axis data matrix. The expression is:

[0100]

[0101] in, It refers to the number of shafts in the first type of shaft. This is the total number of sampling points for each axis. Indicating the first type of axis The first axis corresponds to the current data to be segmented by the sample. One sampling point; The matrix dimension is ;

[0102] S52, Set the length of the sliding window and sliding step size The sliding window is slid according to the set length. For each slide, the current data of all axes in the first type of axis within the same sliding window are extracted to form a sliding window sample. The expression is:

[0103]

[0104] in, Indicates the starting point index of the window is Sliding window samples, This is the index of the starting point of the window.

[0105] S53, Based on sliding step length By continuously sliding the window to a set length, multiple sliding window samples are generated from the original multi-axis data matrix, resulting in a three-dimensional dataset. , is represented as:

[0106]

[0107] in, Indicates co-generation There are 1 sliding window sample, and the size of each sliding window sample is 1. .

[0108] For example, if =4, =2, and the following are two sliding window samples:

[0109] ,

[0110] The determination of the sliding window length and sliding step size is as follows: First, define one cycle using the spindle current data to be sampled; calculate the number of sampling points A corresponding to one cycle under the minimum spindle speed process parameters, and take the smallest power of 2 greater than A as the sliding window length; take the sliding step size as more than half of the sliding window length, so that the window overlaps by about 0%-50%, which covers the complete process features and controls the computational redundancy.

[0111] In the above process, one cycle is defined as the complete engagement and disengagement of the calibrated tool or one full rotation of the spindle. The sliding step size is set to more than half the length of the sliding window, and can be 0.5 to 1 times the length of the sliding window.

[0112] S6. Extract multidimensional time-frequency domain features from the sliding window samples of the three-dimensional dataset under different processing technologies to construct a process feature set. The multidimensional time-frequency domain features include RMS value, peak value, variance, skewness, kurtosis, peak factor, main frequency amplitude, and spectral entropy. Among these, RMS value, peak value, variance, skewness, kurtosis, and peak factor are time domain features, while main frequency amplitude and spectral entropy are frequency domain features.

[0113] S7. Principal component analysis is used to perform nonlinear feature dimensionality reduction on the process feature set, and a multi-class process identification model is constructed by combining a support vector machine with a grid-optimized radial basis function kernel.

[0114] Furthermore, S7 specifically includes:

[0115] S71. Standardize each feature in the process feature set to obtain the standardized feature matrix;

[0116] S72. Principal component analysis is used on the standardized features, with the goal of explaining more than 95% of the cumulative variance. Singular value decomposition is used to calculate the principal components and determine the number of principal components. The standardized feature matrix is ​​then projected onto the principal component space to generate a dimensionality-reduced feature matrix.

[0117] S73. Based on the dimensionality reduction feature matrix, random sampling is performed according to a preset ratio to divide it into a training set and a test set, while maintaining a balanced proportion of each processing technology category.

[0118] S74. Initialize the support vector machine model, select the radial basis function kernel RBF, set the regularization parameter and kernel coefficient, and enable the probability estimation function.

[0119] S75. Use the training set to fit the support vector machine model, optimize the hyperplane to maximize the inter-class margin, and obtain a multi-class process recognition model; use the test set to evaluate the performance of the multi-class process recognition model and verify its classification ability.

[0120] According to a second aspect of the present invention, a CNC machining process identification system based on multi-axis state signals is provided, comprising modules of any of the methods described above. Specifically, it includes: a hardware configuration module, used to, according to the CNC machine tool machining process type, respectively mount current sensors onto the U-phase lines of the servo drive devices of the spindle and the required feed axis within the CNC machine tool electrical cabinet, and connect each current sensor to a data acquisition card for data acquisition; a machining preparation module, used to perform standardized clamping and positioning of the workpiece before each formal machining operation of the CNC machine tool; subsequently, to execute a standardized pre-run program on the CNC machine tool; a multi-axis asynchronous acquisition module, used to treat the spindle with the current sensors and the required feed axis as first-type axes, and for the first-type axes, under different types of machining processes, adopt a condition-triggered asynchronous acquisition method to obtain multi-axis current data; and introduce a timestamp alignment algorithm to achieve temporal consistency of the multi-axis current data, obtaining timestamp-aligned multi-axis current data; data... The reconstruction module combines empirical mode decomposition and wavelet threshold denoising algorithms to construct an adaptive selective denoising preprocessing framework to make decisions on the current data of each axis in the timestamp-aligned multi-axis current data and the reconstructed current data, and uses the current data of each axis determined by the decision as the current data to be segmented as samples. The feature synchronous extraction module is used to obtain a three-dimensional dataset based on the current data to be segmented using multi-axis synchronous sliding window technology. The feature set construction module is used to extract multi-dimensional time-frequency domain features from the current data of each axis in the sliding window samples of the three-dimensional dataset under different types of processing technology to construct a process feature set. The process recognition module uses principal component analysis to perform nonlinear feature dimensionality reduction on the process feature set and combines it with a support vector machine with grid-optimized radial basis function kernels to construct a multi-class process recognition model.

[0121] Example 2: The optional implementation process of the present invention is described below with reference to experimental data:

[0122] I. Implementation Environment

[0123] This embodiment is implemented in the following environment:

[0124] Hardware environment: A computer equipped with an Intel Core i7-9700 processor and 16GB of memory, and a four-axis CNC milling machine XD-40A (X-axis, Y-axis, Z-axis, spindle). The low-speed range of this CNC milling machine is 1-200 mm / min, and the medium-speed range is 200-2000 mm / min.

[0125] Software environment: Python 3.11, LabVIEW 2021.

[0126] Data acquisition equipment: one NI 6003 acquisition card and three current sensors.

[0127] II. Implementation Steps

[0128] In this embodiment, four typical machining processes were selected for testing: X-shaped milling, Y-shaped milling, circular arc machining, and oblique line machining. Data details are shown in Table 1 below (for example, the process parameters in Table 1 are described as follows: 1000-120-0.5, spindle speed 1000rpm, feed rate 120mm / min, depth of cut 0.5mm, and so on).

[0129] Table 1 Data for different processing technology types

[0130]

[0131] According to the method of the present invention, the optional process of this embodiment is specifically described as follows:

[0132] 1) Based on the CNC machine tool machining process type in Table 1 above, place three current sensors on the X-axis, Y-axis and U-phase lines of the spindle servo drive device in the electrical cabinet of the CNC machine tool, and connect each current sensor to the NI 6003 data acquisition card for data acquisition.

[0133] 2) Workpiece clamping and positioning stage: Based on the material properties and geometry of the nylon workpiece, a fixture with a contact area of ​​more than 60% between the workpiece and the fixture is selected as the main solution. The fixture is used to apply a predetermined clamping force (approximately 800–1000N) under controlled pressure to prevent the nylon workpiece from being deformed under pressure and to ensure that the clamping force is evenly distributed. After clamping, a dial indicator is used to check the key positioning surfaces of the workpiece to verify that the clamping error is controlled within ±0.02mm, thereby ensuring the positioning accuracy and stability of subsequent processing.

[0134] Specifically, in the actual experiment, various clamping devices were used for comparison, including mechanical flat-jaw vises, V-block clamps, pressure plates, and hydraulic vises. Measurements and comparisons revealed that the hydraulic vise, with the addition of soft jaw pads, could form an effective contact area of ​​over 60% with the nylon workpiece, and its clamping force distribution was uniform, with better repeatability than other clamping devices. Therefore, the hydraulic vise, with the largest contact area and strongest clamping stability, was ultimately selected as the primary clamping solution. At the preset position, a predetermined clamping force (approximately 800–1000 N) was applied using the hydraulic vise under controlled pressure. A dial indicator was then used to check the key positioning surfaces of the workpiece, verifying that the clamping error was within ±0.02 mm, indicating that the current preset position met the processing conditions. It should be noted that if the clamping error did not meet the "within ±0.02 mm" standard after verification, the workpiece position was readjusted, and the predetermined clamping force was applied until it met the preset standard. The preset standard is an error within ±0.02 mm.

[0135] 3) The CNC machine tool executes the standardized pre-running procedure:

[0136] Spindle idling and preheating stage: The spindle is run under no-load self-test to allow the spindle motor to reach a stable operating temperature and eliminate electrical system abnormalities.

[0137] Low-speed no-load reciprocating motion stage of each feed axis: Under no-load conditions, perform a low-speed reciprocating motion self-check on each feed axis to eliminate the backlash of each feed axis;

[0138] Medium-speed no-load reciprocating motion stage of each feed axis: Under no-load conditions, each feed axis performs medium-speed reciprocating motion, so that the corresponding servo motor completes the medium-speed reciprocating motion self-check and achieves parameter stability.

[0139] Return to machine origin operation phase: Reset each feed to the reference initial position;

[0140] Self-inspection stage at the beginning of machining: By executing a unified and standardized operation process for spindle lifting and lowering, the consistency of the state at the beginning of machining is ensured.

[0141] 4) The spindle with the current sensor and the required feed axis are designated as the first type of axis. For the first type of axis, under different types of machining processes, a non-synchronous acquisition method based on working condition triggering is adopted to obtain multi-axis current data, namely the current data of the X-axis, Y-axis, and spindle. The current data of each axis consists of 1,996,800 sampling points. A timestamp alignment algorithm is introduced to achieve consistency in the timing of the multi-axis current data. Specifically: the spindle is selected as the reference signal from the multi-axis current data; X-shaped milling: the spindle is used as the reference signal, the X-axis is timestamped, and the Y-axis is not timestamped; Y-shaped milling: the spindle is used as the reference signal, the Y-axis is timestamped, and the X-axis is not timestamped; Circular milling: the spindle is used as the reference signal, and timestamps are introduced for both the X-axis and Y-axis; Oblique milling: the spindle is used as the reference signal, and timestamps are introduced for both the X-axis and Y-axis.

[0142] 5) Perform Empirical Mode Decomposition (EMD) on the current data of each axis to adaptively decompose the nonlinear, non-stationary current signal into a series of Intrinsic Mode Functions (IMFs). If any IMF component satisfies either the frequency domain energy distribution characteristics or statistical characteristics criteria, it is identified as a noise IMF component. The frequency domain energy distribution characteristics criteria are: 25% to 50% of the sampling frequency is defined as the high-frequency range, corresponding to the high-frequency energy interval of 6400Hz-12800Hz; when the high-frequency energy of a certain IMF component accounts for more than 0.6% of its total energy, that IMF is a noise IMF component; if the high-frequency energy of the first IMF accounts for more than 0.4%, it is also considered a noise IMF component. The statistical characteristics criteria are: when the variance of the first two IMF components is less than 10% of the total variance of the original signal, it is considered a noise IMF component. Noise IMF components that satisfy either of the above criteria are uniformly marked as noise reduction targets. Then, wavelet threshold denoising algorithm is applied only to the identified noisy IMF components, while keeping the remaining useful IMF components intact. Finally, the energy retention rate test (threshold set to 0.3) is used to ensure the quality of the reconstructed signal. When the energy retention rate is lower than the preset threshold, the energy loss is considered too large, and the original current data before reconstruction is determined to be the current data to be segmented. Otherwise, the reconstructed current data is determined to be the current data to be segmented. The flowchart is as follows. Figure 3 As shown.

[0143] 6) Based on the current data to be segmented, a three-dimensional dataset is obtained using multi-axis synchronous sliding window technology. Specifically, the dimensions of the obtained three-dimensional dataset are 975×3×2048 for each type of machining process. First, the spindle current data to be segmented is used to define the complete engagement and disengagement of the calibration tool as one cycle. The number of sampling points (approximately 1536 points) corresponding to one cycle under the lowest spindle speed process parameters is calculated (the process parameters are the 13 types mentioned in Table 1). To ensure that each window covers at least one complete cycle, and to make the Fast Fourier Transform (FFT) operation run faster using a power of 2 length, in this embodiment of the invention, the window length is 2048 sampling points. The sliding step size is taken as one time the sliding window length (2048 sampling points) to ensure 0% window overlap, thus covering the complete process features while controlling computational redundancy. The minimum speed process parameters in Table 1 are 1000-120-0.5, 1000-200-0.5, and 1000-320-0.5. You can choose any one of them.

[0144] 7) Extract multi-dimensional time-frequency domain features from the current data of each axis in the sliding window samples of the 3D dataset (i.e., extract the time-domain and frequency-domain features of each axis under each sliding window sample) to construct a process feature set; among which, the multi-dimensional time-frequency domain features include RMS value, peak value, variance, skewness, kurtosis, peak factor, main frequency amplitude, and spectral entropy. The specific calculation method is as follows:

[0145] Valid values:

[0146] in, For the first Current data values ​​at each sampling point This represents the total number of sampling points.

[0147] Peak value:

[0148] in, This represents the absolute value of the current data. This indicates taking the maximum value.

[0149] variance:

[0150] in, This represents the average current data.

[0151] Skewness:

[0152] in, This represents the signal mean.

[0153] Peak factor:

[0154] The numerator represents the peak value, and the denominator represents the effective value.

[0155] kurtosis:

[0156] in, This represents the signal mean.

[0157] Main frequency amplitude: ;

[0158] in, Frequency components amplitude, This indicates taking the maximum value.

[0159] Spectral entropy is calculated through the following steps:

[0160] (1) Calculate the normalized power spectral density:

[0161] ;

[0162] (2) Calculate the spectral entropy:

[0163]

[0164] Among them, if ,but The value is 0.

[0165] Based on the above, by applying the above methods to the four processing techniques, 975 24-dimensional samples can be obtained respectively, with 8 features extracted from each of the 24 dimensions (3 axes).

[0166] 8) Principal component analysis is used to perform nonlinear feature dimensionality reduction on the process feature set, and a multi-class process identification model is constructed by combining a support vector machine with a grid-optimized radial basis function kernel, so as to realize the automatic classification and identification of various processing technologies.

[0167] S81. Standardize each feature in the process feature set using StandardScaler to scale each feature to zero mean and unit variance, resulting in a standardized feature matrix. For example, the matrix has the number of rows as the number of samples in the feature set and the number of columns as the number of feature dimensions. In this embodiment of the invention, under the four processing technologies, the number of samples in the feature set is 3900 and the feature dimension is 24.

[0168] S82. Principal component analysis (PCA) is used on the standardized features with the goal of explaining more than 95% of the cumulative variance. Singular value decomposition is used to calculate the principal components and determine the number of principal components. The standardized feature matrix is ​​projected onto the principal component space to generate a dimensionality-reduced feature matrix (3900×10).

[0169] S83. Based on the dimensionality reduction feature matrix, random sampling is performed at a ratio of 7:3 to divide the data into a training set and a test set (i.e., the training set contains 2730 samples from four processes, and the test set contains 1170 samples), while maintaining a balanced proportion of each processing process category.

[0170] S84. Initialize the Support Vector Machine (SVM) model, select the Radial Basis Function (RBF) kernel, and set the regularization parameters. kernel coefficient And enable probability estimation function;

[0171] S85. Use the training set to fit the support vector machine model, optimize the hyperplane to maximize the inter-class margin, and obtain a multi-class process recognition model; use the test set to evaluate the performance of the multi-class process recognition model and verify its classification ability.

[0172] Using the above-described method of the present invention, after testing with a test set, the multi-process identification model of the present invention meets the classification capability, and the multi-process identification model is saved. Subsequently, asynchronous acquisition is performed on the data under new unknown processing technology to obtain multi-axis current data; then, multi-axis current data with timestamp alignment is obtained, and then current data to be segmented is obtained; further, multi-axis synchronous sliding window technology is used to obtain a three-dimensional dataset, and dimensionality reduction features are extracted to construct a feature set of the process to be identified; based on the saved multi-process identification model, the processing technology can be identified.

[0173] III. Implementation Results

[0174] This embodiment uses a combination of principal component analysis (PCA) dimensionality reduction and support vector machine (SVM) classification to identify four process types (X-shaped milling, Y-shaped milling, circular arc machining, and oblique line machining), as shown in Table 2. The accuracy, recall, and F1 score are used to evaluate each process.

[0175] Table 2. Evaluation Indicators for Process Identification

[0176]

[0177] As shown in Table 2, the overall classification accuracy reached 97.52%. The model demonstrated exceptional performance in the classification tasks of X-shaped and Y-shaped milling, with precision and recall both approaching 1.0, reflecting its strong robustness to linear process features. It also maintained a certain level of recognition ability in the classification of arc and oblique machining, fully demonstrating the algorithm's stability and generalization ability when handling diverse process signals.

[0178] Furthermore, by conducting real-time prediction experiments using the aforementioned stored multi-process identification models, the prediction results of randomly selected test samples are as follows: Figure 4As shown, the model can correctly identify the process type with a high confidence probability >0.95 in most cases, and the predicted probability distribution is highly consistent with the true label. This verifies that the method can effectively adapt to the dynamically changing process classification needs in industrial environments and has high engineering application value. The implementation is as follows: First, multi-axis current data is acquired asynchronously; then, timestamp-aligned multi-axis current data is obtained; next, an adaptive selective denoising preprocessing framework combining empirical mode decomposition and wavelet threshold denoising algorithms is constructed to make decisions on the current data of each axis in the timestamp-aligned multi-axis current data and the reconstructed current data to determine the current data to be segmented; further, multi-axis synchronous sliding window technology is used to obtain a three-dimensional dataset, and dimensionality reduction features are extracted to construct a feature set of the process to be identified; finally, the process is identified based on the saved multi-class process identification models. The above-mentioned data is stored in a streamlined manner and can be easily migrated to an edge computing platform in the industrial field to achieve millisecond-level inference. This solution has low computational overhead and simple maintenance, which can significantly reduce the cost of manual judgment and can quickly screen steady-state data under the same process, providing reliable support for subsequent fault diagnosis and performance evaluation of CNC machine tools. This fully demonstrates the engineering application value of this method in intelligent manufacturing scenarios.

[0179] The specific embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A method for identifying CNC machining processes based on multi-axis state signals, characterized in that, include: S1. Based on the CNC machine tool processing technology type, place the current sensors on the U-phase lines of the servo drive devices of the spindle and the required feed axis in the electrical cabinet of the CNC machine tool, and connect each current sensor to the data acquisition card. S2. Before each formal machining operation of a workpiece on the CNC machine tool, the workpiece is clamped and positioned in a standardized manner; then the CNC machine tool is executed a standardized pre-run program. S3. Take the spindle with the current sensor installed and the required feed axis as the first type of axis. For the first type of axis, under different types of machining processes, adopt the asynchronous acquisition method based on working condition triggering to obtain multi-axis current data respectively. Introduction The timestamp alignment algorithm achieves the consistency of multi-axis current data in terms of timing, and obtains multi-axis current data after timestamp alignment. S4. An adaptive selective noise reduction preprocessing framework is constructed by combining empirical mode decomposition and wavelet threshold noise reduction algorithm to make decisions on the current data of each axis in the timestamp-aligned multi-axis current data and the reconstructed current data, and the current data of each axis determined by the decision is used as the current data to be segmented as samples. S5. Based on the current data to be segmented, a three-dimensional dataset is obtained using multi-axis synchronous sliding window technology; S6. Extract multi-dimensional time-frequency domain features from the sliding window samples of the three-dimensional dataset under different types of processing technology, and construct a process feature set; S7. Principal component analysis is used to perform nonlinear feature dimensionality reduction on the process feature set, and a multi-class process identification model is constructed by combining a support vector machine with a grid-optimized radial basis function kernel. Specifically, S5 is: S51. Based on the current data to be segmented from the sample, establish a multi-axis data matrix. The expression is: ； in, It refers to the number of shafts in the first type of shaft. This is the total number of sampling points for each axis. Indicating the first type of axis The first axis corresponds to the current data to be segmented by the sample. One sampling point; The matrix dimension is ; S52, Set the length of the sliding window and sliding step size The sliding window is slid according to the set length. For each slide, the current data of all axes in the first type of axis within the same sliding window are extracted to form a sliding window sample. The expression is: ； in, Indicates the starting point index of the window is Sliding window samples, The index of the starting point of the window; S53, Based on sliding step length By continuously sliding the window to a set length, multiple sliding window samples are generated from the original multi-axis data matrix, resulting in a three-dimensional dataset. , represented as: ； in, Indicates co-generation There are 1 sliding window sample, and the size of each sliding window sample is 1. .

2. The CNC machining process identification method based on multi-axis state signals according to claim 1, characterized in that, S2 specifically includes: S21. Before each formal machining of a workpiece on a CNC machine tool, the workpiece shall be clamped and positioned in a standardized manner using standard fixtures. S22. Enable the CNC machine tool to execute a standardized pre-run program: Spindle idling and preheating stage: The spindle is run under no-load self-test to allow the spindle motor to reach a stable operating temperature and eliminate electrical system abnormalities. Low-speed no-load reciprocating motion stage of each feed axis: Under no-load conditions, perform a low-speed reciprocating motion self-check on each feed axis; Medium-speed no-load reciprocating motion stage of each feed axis: Under no-load conditions, each feed axis performs medium-speed reciprocating motion, so that the corresponding servo motor completes the medium-speed reciprocating motion self-check and achieves parameter stability. Return to machine origin operation phase: Reset each feed to the reference initial position; Self-inspection stage at the beginning of machining: By executing a unified and standardized operation process for spindle lifting and lowering, the consistency of the state at the beginning of machining is ensured.

3. The CNC machining process identification method based on multi-axis state signals according to claim 2, characterized in that, The standardized workpiece clamping and positioning specifically includes: selecting a clamp that ensures the contact area between the workpiece and the clamp reaches more than 60% based on the workpiece's material properties and geometry; tightening the clamp at a preset position using a predetermined clamping force; and ensuring that the clamping error detected by a coordinate measuring machine or indicator meets the preset standard.

4. The CNC machining process identification method based on multi-axis state signals according to claim 1, characterized in that, The method of obtaining multi-axis current data by using a non-synchronous acquisition method based on working condition triggering is as follows: the end of the unified and standardized operation process of the spindle lifting and lowering the tool is used as the trigger condition for starting the current sensor; after the trigger condition is met, the current sensors of the spindle and the required feed axis are started to acquire the current signal with the same number of sampling points, thereby obtaining multi-axis current data.

5. The CNC machining process identification method based on multi-axis state signals according to claim 1, characterized in that, Specifically, S4 is: Empirical mode decomposition is performed on the current data of each axis in the timestamp-aligned multi-axis current data to obtain multiple IMF components; If any IMF component satisfies any judgment criterion based on frequency domain energy distribution characteristics or statistical properties, it is identified as a noise IMF component. The wavelet threshold denoising algorithm is applied only to the identified noisy IMF components, while keeping the other IMF components unchanged, to obtain the reconstructed current data. The energy retention rate of the reconstructed current data is examined, and a decision is made based on the energy retention rate: if the energy retention rate is lower than a preset threshold, the original current data before reconstruction is determined to be the current data to be segmented; otherwise, the reconstructed current data is determined to be the current data to be segmented.

6. The CNC machining process identification method based on multi-axis state signals according to claim 1, characterized in that, The determination of the sliding window length and sliding step size is as follows: First, define one cycle using the current data of the spindle sample to be divided; calculate the number of sampling points A corresponding to one cycle under the minimum spindle speed process parameters, and take the smallest power of 2 greater than A as the sliding window length; take the sliding step size as more than half of the sliding window length.

7. The CNC machining process identification method based on multi-axis state signals according to claim 1, characterized in that, Specifically, S7 is: S71. Standardize each feature in the process feature set to obtain the standardized feature matrix; S72. Principal component analysis is used on the standardized features, with the goal of explaining more than 95% of the cumulative variance. Singular value decomposition is used to calculate the principal components and determine the number of principal components. The standardized feature matrix is ​​then projected onto the principal component space to generate a dimensionality-reduced feature matrix. S73. Based on the dimensionality reduction feature matrix, random sampling is performed according to a preset ratio to divide it into a training set and a test set, while maintaining a balanced proportion of each processing technology category. S74. Initialize the support vector machine model, select the radial basis function kernel RBF, set the regularization parameter and kernel coefficient, and enable the probability estimation function. S75. Use the training set to fit the support vector machine model, optimize the hyperplane to maximize the inter-class margin, and obtain a multi-class process recognition model; use the test set to evaluate the performance of the multi-class process recognition model and verify its classification ability.

8. A CNC machining process recognition system based on multi-axis status signals, characterized in that, The module includes the CNC machining process identification method based on multi-axis status signals as described in any one of claims 1-7.

Images