A method and system for predicting surface roughness of an inner profile hole based on a grinding force signal

Abstract

The application discloses a kind of inner circle profiled hole grinding surface roughness prediction method and system based on grinding force signal, and it is related to the field of precision grinding processing and intelligent manufacturing technology, this method mainly includes: to three-way grinding force signal is de-trended, filtering and angle domain etc. Angle resampling is obtained Angle domain force signal, the geometric parameter of each angle position is calculated, according to angle window extraction force signal Statistics features, dynamic characteristics and fusion features, utilize hybrid prediction model to predict the surface roughness distribution of non-circular curve hole full profile and identify roughness abnormal area.The grinding surface roughness prediction method and system based on grinding force signal provided in the application can improve the accuracy and generalization of non-circular curve hole full profile surface roughness online distribution prediction.

Description

Technical Field

[0001] This invention relates to the fields of precision grinding and intelligent manufacturing technology, and more specifically, to a method and system for predicting the surface roughness of grinding inner circular irregular holes based on grinding force signals. Background Technology

[0002] Non-circular curve holes refer to hole types whose inner contours are composed of non-standard circular curves, including elliptical holes, parabolic holes, hyperbolic holes, involute holes, and B-spline curve holes. Due to their unique advantages in fluid dynamics control, weight reduction, and stress distribution optimization, these hole types are widely used in key components such as weight reduction holes for aerospace engine turbine disks, cooling holes for casings, atomizing holes for fuel nozzles, irregularly shaped cavities in precision molds, and internal cavities of medical devices. Taking aerospace engine turbine disks as an example, to meet weight reduction and cooling requirements, dozens of elliptical weight reduction holes and irregularly shaped cooling holes are distributed on the disk body. Their machining quality directly affects the fatigue life and reliability of the turbine disk. However, the grinding of non-circular curve holes presents the following technical challenges: (I) Dynamic Time-Varying of Grinding Force: The essential characteristic that distinguishes non-circular curved holes from traditional circular holes lies in the continuous variation of their contour radius of curvature along the circumference. Taking elliptical holes as an example, the ratio of the radius of curvature at the major axis endpoint to that at the minor axis endpoint can reach 2-3 times. This variation in radius of curvature causes periodic fluctuations in parameters such as the grinding contact arc length, the entry and exit angles, and the equivalent grinding wheel diameter. Experiments show that the peak and valley values ​​of the normal force Fn during elliptical hole grinding can differ by 30%-50%. Traditional constant-parameter grinding methods cannot adapt to this dynamic variation, easily leading to uneven surface quality and excessive machining accuracy.

[0003] (ii) Differences in the profile distribution of surface roughness: Due to the dynamic time-varying nature of grinding force, the surface roughness at different angular positions of non-circular curved holes exhibits significant differences in distribution. Measured data shows that the surface roughness difference between the major axis endpoint (the point of maximum curvature) and the minor axis endpoint (the point of minimum curvature) of an elliptical hole can reach 30%-50%. Traditional offline sampling inspection methods typically only measure 1-3 points within the hole, failing to reflect the overall profile quality distribution and unable to identify weak areas, thus posing a serious risk of missed inspections.

[0004] (iii) Lack of online inspection methods: The following problems exist in measuring the inner surface of non-circular curved holes using a contact roughness meter: (1) It is difficult to center the measuring probe on the surface with variable curvature, resulting in poor measurement repeatability; (2) The measurement process is prone to scratching the machined surface, which is particularly unfavorable for precision-machined aerospace parts; (3) The efficiency of point-by-point measurement is extremely low, making it impossible to achieve batch online inspection; (4) It is impossible to obtain quality information in real time during the grinding process, making it difficult to form a closed-loop control.

[0005] (iv) Limitations of existing prediction methods: Scholars at home and abroad have conducted a lot of research on the prediction of grinding surface roughness. The main methods can be divided into three categories: (1) Methods based on empirical formulas: such as the Snoeys-Brown model, Malkin model, etc., which have established empirical relationships between grinding parameters and surface roughness. This type of method assumes constant grinding conditions and cannot handle the dynamic effects caused by changes in the curvature of non-circular curved holes. (2) Data-driven method based on a single sensor: signals are collected using vibration sensors, acoustic emission sensors or power sensors, and predictions are made using models such as neural networks and support vector machines. However, this type of method is difficult to interpret signals under variable curvature conditions and lacks angular correlation with the contour position. (3) Machine vision-based method: features are extracted by collecting images of the grinding area using a CCD camera. This method is severely affected by coolant and grinding debris and is difficult to install in the narrow space of internal cylindrical grinding.

[0006] In summary, the existing technology has the following technical defects: (1) lacks a dedicated prediction method for the curvature characteristics of non-circular curve holes; (2) cannot achieve online prediction of the roughness distribution of the entire contour; (3) the pure data-driven model has a large demand for training samples and insufficient generalization ability; (4) lacks a quantitative compensation mechanism for the influence of curvature changes on roughness.

[0007] Overcoming the shortcomings of existing technologies and using grinding force signals to predict the surface roughness distribution along the contour of non-circular curved holes in real time, thereby improving the accuracy and generalization of online prediction of the surface roughness distribution of the entire contour of non-circular curved holes, is of great significance for achieving high-quality and high-efficiency grinding of non-circular curved holes. Summary of the Invention

[0008] The purpose of this invention is to provide a method and system for predicting the surface roughness of grinding inner circular irregular holes based on grinding force signals, which can improve the accuracy and generalization of online prediction of the surface roughness distribution of the entire contour of non-circular curved holes.

[0009] This invention provides a method for predicting the surface roughness of internal circular irregular holes based on grinding force signals, comprising the following steps: S1: The triaxial grinding force signal in the non-circular curve hole grinding process is preprocessed to obtain the preprocessed triaxial grinding force signal. The preprocessed force signal is then resampled in the angle domain according to the spindle rotation angle signal to convert the time domain triaxial grinding force signal into the angle domain triaxial grinding force signal. S2: Calculate the geometric parameters of the non-circular curve hole at each angular position, including the radius of curvature, equivalent grinding wheel diameter, and grinding contact arc length; S3: The triaxial grinding force signal in the angle domain is segmented according to the angle window of the preset width, and the statistical features, dynamic features, and fusion features of the statistical features and the geometric parameters of each angle window are extracted. S4: Construct a hybrid prediction model including a physical model layer and a neural network model layer. The physical model layer is used to obtain a prior physical prediction including a curvature-sensitive compensation term based on the theoretical relationship between grinding force and surface roughness. The neural network model layer is used to obtain the roughness prediction deviation based on the statistical features, dynamic features and fusion features. S5: Based on the statistical features, dynamic features, and fusion features, the roughness prediction distribution of the entire contour of the non-circular curve hole is obtained using the hybrid prediction model. The angle region corresponding to the preset threshold where the roughness in the roughness prediction distribution exceeds the average roughness of the entire contour is taken as the roughness anomaly region.

[0010] Furthermore, the non-circular curve hole includes at least one of elliptical hole, parabolic hole, hyperbolic hole, involute hole, and B-spline curve hole; the triaxial grinding force signal includes normal force, tangential force, and axial force.

[0011] Furthermore, the preprocessing includes removing trend terms to eliminate zero-point drift and low-pass filtering.

[0012] Furthermore, the formula for calculating the geometric parameters is as follows:

[0013]

[0014]

[0015] in, The radius of curvature; The equivalent grinding wheel diameter; This is the grinding contact arc length; For polar angle In terms of direction, the distance from the pole to the contour point of the inner wall of the hole; and They are respectively First and second derivatives with respect to θ; The diameter of the grinding wheel; This refers to the grinding depth.

[0016] Furthermore, the statistical features include the mean normal force, mean tangential force, mean axial force, root mean square of normal force, root mean square of tangential force, root mean square of axial force, and force ratio; the dynamic features include the coefficient of variation, the peak frequency and peak amplitude of the three-axis grinding force signal; the fusion features of the statistical features and the geometric parameters include the unit normal force and grinding specific energy.

[0017] Furthermore, the expression for the hybrid prediction output of the hybrid prediction model is:

[0018] in, For the hybrid prediction model for the first Mixed prediction output from multiple angle windows; For the first Prior physics predictions for each angle window; For the first Roughness prediction bias for each angle window.

[0019] Furthermore, the calculation formula for the prior physical prediction is as follows:

[0020] in: For the first Prior physics predictions for each angle window; Material correlation coefficient; For the first The average tangential force of each angle window; The linear velocity of the grinding wheel; For the first Workpiece rotation speed at each angle window; This refers to the grinding depth. The curvature sensitivity coefficient; For the first The radius of curvature of each angle window; Use the radius of curvature as a reference. This refers to the basic surface roughness of the workpiece.

[0021] Furthermore, the neural network model layer is a deep neural network, and the structured deep neural network includes an input layer, a first hidden layer, a second hidden layer, a third hidden layer, and an output layer; the input layer has a dimension of 8-15 and is used to receive the statistical features, dynamic features, and fused features; the first hidden layer has 64 nodes, and each node includes the activation function ReLU and batch normalization; the second hidden layer has 32 nodes, and each node includes the activation function ReLU and batch normalization; the third hidden layer has 16 nodes, and each node includes the activation function ReLU; the output layer is a single node, including linear activation, and is used to output the roughness prediction bias.

[0022] Furthermore, the above-mentioned method for predicting the surface roughness of internal circular irregular holes based on grinding force signals also includes: For each preset number of workpieces processed, one workpiece is randomly selected and its roughness is measured offline at a preset number of characteristic angle positions to obtain the measured value. The characteristic angle positions include curvature extreme points and curvature inflection points. The difference between the measured value and the predicted value of the hybrid prediction model is used as feedback. The weights of the neural network model layer of the hybrid prediction model are updated using incremental learning. Kalman filtering is used to correct the material correlation coefficient, curvature sensitivity coefficient and workpiece basic roughness of the physical model layer of the hybrid prediction model online, so as to obtain the updated hybrid prediction model.

[0023] This invention also provides a surface roughness prediction system for grinding internal circular irregular holes based on grinding force signals, comprising the following modules: The signal preprocessing module is used to: preprocess the triaxial grinding force signal in the non-circular curve hole grinding process to obtain the preprocessed triaxial grinding force signal, and perform angular domain equal-angle resampling on the preprocessed force signal according to the spindle rotation angle signal to convert the time domain triaxial grinding force signal into the angular domain triaxial grinding force signal. The geometric parameter calculation module is used to calculate the geometric parameters of non-circular curve holes at various angular positions, including the radius of curvature, equivalent grinding wheel diameter, and grinding contact arc length. The feature extraction module is used to: segment the triaxial grinding force signal of the angle domain according to the angle window of the preset width, and extract the statistical features, dynamic features and the fusion features of the statistical features and the geometric parameters of each angle window; The hybrid prediction model module is used to: construct a hybrid prediction model including a physical model layer and a neural network model layer. The physical model layer is used to obtain a prior physical prediction including a curvature-sensitive compensation term based on the theoretical relationship between grinding force and surface roughness. The neural network model layer is used to obtain the roughness prediction deviation based on the statistical features, dynamic features, and fusion features. The abnormal region identification module is used to: obtain the roughness prediction distribution of the entire contour of the non-circular curve hole based on the statistical features, dynamic features and fusion features using the hybrid prediction model, and identify the angle region corresponding to the roughness exceeding the preset threshold of the roughness of the entire contour in the roughness prediction distribution as the roughness abnormal region.

[0024] The method and system for predicting the surface roughness of grinding internal circular irregular holes based on grinding force signals provided by this invention have the following beneficial effects: This invention addresses the shortcomings of existing technologies, such as uneven surface quality distribution and difficulty in online detection caused by the continuous change of the radius of curvature along the contour during the grinding of non-circular curve holes (including elliptical holes, parabolic holes, hyperbolic holes, involute holes, and B-spline curve holes). It provides online monitoring of the surface quality during the grinding process of non-circular curve holes, and specifically predicts the surface roughness of non-circular curve holes based on the angular domain analysis of the grinding force signal.

[0025] First, this invention uses a force gauge to collect triaxial grinding force signals during the grinding process of non-circular curved holes, and simultaneously collects spindle rotation angle signals. The force signals are then de-trended, filtered, and resampled in the angle domain to obtain angle domain force signals. Second, geometric parameters such as the radius of curvature, equivalent grinding wheel diameter, and grinding contact arc length at each angle position are calculated. Then, statistical features, dynamic features, and fusion features of the force signals are extracted according to the angle window. Furthermore, a hybrid prediction model is constructed, including a physical model layer with embedded curvature-sensitive compensation terms and a deep neural network deviation prediction layer. Finally, the surface roughness distribution of the entire contour of the non-circular curved hole is output, and roughness anomaly regions are identified.

[0026] This invention enables online prediction of the surface roughness distribution of non-circular curved holes across their entire contour. Through a physical-data hybrid driving method, it improves prediction accuracy and generalization ability, achieving an average relative prediction error of less than 4%. It is applicable to the precision grinding of non-circular curved holes in difficult-to-machine materials such as aerospace high-temperature alloys and titanium alloys. Specifically, in the field of precision grinding technology, this invention is applicable to the internal grinding of non-circular curved holes (including elliptical holes, parabolic holes, hyperbolic holes, involute holes, and B-spline curve holes) in key components such as aerospace engine turbine disks, casings, and fuel nozzles, using high-temperature alloys and titanium alloys. This invention enables online prediction of surface roughness and full-contour quality assessment during the grinding process. In the field of online measurement technology for the grinding process, this invention utilizes the grinding force signal collected by a force gauge as an information source, and achieves indirect measurement and prediction of surface roughness through signal processing and feature extraction. In the field of intelligent manufacturing and process monitoring, this invention constructs a hybrid prediction model (physical model + neural network) to achieve online assessment of grinding status and identification of abnormal areas, providing a decision-making basis for adaptive adjustment of process parameters. In the field of materials processing, this invention is particularly suitable for grinding non-circular curve holes in difficult-to-machine materials (such as high-temperature alloys GH4169, Inconel 718, titanium alloys TC4, etc.), which are widely used in key components such as turbine disks, casings, and fuel nozzles of aerospace engines. In addition, this invention is also applicable to the interdisciplinary fields of mechanical processing and information science, integrating grinding mechanics, signal processing, machine learning (deep neural networks), and geometric modeling. Attached Figure Description

[0027] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is a flowchart of the method for predicting the surface roughness of grinding internal circular irregular holes based on grinding force signals provided by the present invention. Figure 2 This is a schematic diagram of the operation steps of the method for predicting the surface roughness of grinding internal circular irregular holes based on grinding force signals provided by the present invention. Figure 3This is a schematic diagram of the non-circular curve hole grinding force signal acquisition system provided by the present invention; Figure 4 This is a schematic diagram of the conversion between time-domain force signals and angle-domain force signals provided by the present invention; Figure 5 The radius of curvature of the elliptical hole profile provided by this invention Schematic diagram of the curve pattern as a function of angle; Figure 6 This is a schematic diagram showing the variation curves of grinding force characteristics at different angle windows provided by the present invention; Figure 7 This is a structural diagram of the hybrid prediction model provided by the present invention; Figure 8 This is a schematic diagram comparing the predicted surface roughness profile of the elliptical hole grinding surface with the measured value provided by the present invention; Figure 9 This is a schematic diagram of the process of generating a heat map of the predicted roughness distribution of the full contour of a parabolic hole provided by the present invention. Figure 10 This is a flowchart of the online model update process provided by the present invention; Figure 11 This is a schematic diagram comparing the prediction accuracy of the method provided by this invention with existing methods. Detailed Implementation

[0028] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0029] Figure 1 A schematic diagram of the method for predicting the surface roughness of an internal circular irregular hole based on grinding force signals, according to this embodiment, is shown. In this embodiment, the method for predicting the surface roughness of an internal circular irregular hole based on grinding force signals includes the following steps: S1: The triaxial grinding force signal in the non-circular curve hole grinding process is preprocessed to obtain the preprocessed triaxial grinding force signal. The preprocessed force signal is then resampled in the angle domain according to the spindle rotation angle signal to convert the time domain triaxial grinding force signal into the angle domain triaxial grinding force signal. In one exemplary embodiment, the non-circular curve hole includes at least one of elliptical hole, parabolic hole, hyperbolic hole, involute hole, and B-spline curve hole; The triaxial grinding force signal includes normal force, tangential force, and axial force; As an exemplary embodiment, the profile of the non-circular curved hole is given by the polar coordinate equation r=f(θ) or the parametric equation form; In one exemplary embodiment, the preprocessing includes removing trend terms to eliminate zero-point drift and low-pass filtering; In one exemplary embodiment, the cutoff frequency of the low-pass filter is 500Hz; the number of sampling points for the angular domain equal-angle resampling is N=360 per revolution. 720 sampling points, with an angular resolution of 0.5°-1°; S2: Calculate the geometric parameters of the non-circular curve hole at each angular position, including the radius of curvature, equivalent grinding wheel diameter, and grinding contact arc length; In one exemplary embodiment, the formula for calculating the geometric parameters is:

[0030]

[0031]

[0032] in, The radius of curvature; The equivalent grinding wheel diameter; This is the grinding contact arc length; For polar angle In terms of direction, the distance from the pole to the contour point of the inner wall of the hole; and They are respectively First and second derivatives with respect to θ; The diameter of the grinding wheel; This refers to the grinding depth. S3: The triaxial grinding force signal in the angle domain is segmented according to the angle window of the preset width, and the statistical features, dynamic features, and fusion features of the statistical features and the geometric parameters of each angle window are extracted. In one exemplary embodiment, the preset width is 5°. 15°; the overlap rate of adjacent angled windows is 50%; In one exemplary embodiment, the statistical features include mean normal force, mean tangential force, mean axial force, root mean square of normal force, root mean square of tangential force, root mean square of axial force, and force ratio. The dynamic characteristics include the coefficient of variation, the peak frequency of the triaxial grinding force signal, and the peak amplitude of the spectrum. The fusion of the statistical features and the geometric parameters includes unit normal force and grinding specific energy; In one exemplary embodiment, the formula for calculating the force ratio is:

[0033] in, Indicates the first The force ratio of each angle window; For the first The average normal force of each angle window; For the first The average tangential force of each angle window; In one exemplary embodiment, the formula for calculating the coefficient of variation is:

[0034] in, For the first The coefficient of variation for each angle window; For the normal force within the i-th angular window The sample standard deviation; For the first The average normal force of each angle window; In one exemplary embodiment, the formula for calculating the unit normal force is:

[0035] in, For the first The unit normal force of a window at each angle; For the first The average normal force of each angle window; For the first The grinding contact arc length of each angle window; In one exemplary embodiment, the formula for calculating the grinding specific energy is:

[0036] in, For the first Grinding specific energy of each angle window; For the first The average tangential force of each angle window; The linear velocity of the grinding wheel; For the first Material removal rate of each angle window; S4: Construct a hybrid prediction model including a physical model layer and a neural network model layer. The physical model layer is used to obtain a prior physical prediction including a curvature-sensitive compensation term based on the theoretical relationship between grinding force and surface roughness. The neural network model layer is used to obtain the roughness prediction deviation based on the statistical features, dynamic features and fusion features. In one exemplary embodiment, the expression for the hybrid prediction output of the hybrid prediction model is:

[0037] in, For the hybrid prediction model for the first Mixed prediction output from multiple angle windows; For the first Prior physics predictions for each angle window; For the first Roughness prediction bias for each angle window; In one exemplary embodiment, the formula for calculating the prior physical prediction is:

[0038] in: For the first Prior physics predictions for each angle window; Material correlation coefficient; For the first The average tangential force of each angle window; The linear velocity of the grinding wheel; For the first Workpiece rotation speed at each angle window; This refers to the grinding depth. The curvature sensitivity coefficient; For the first The radius of curvature of each angle window; Used as the reference radius of curvature; For the basic surface roughness of the workpiece; In one exemplary embodiment, the neural network model layer is a deep neural network, which includes an input layer, a first hidden layer, a second hidden layer, a third hidden layer, and an output layer. The input layer has a dimension of 8-15 and is used to receive the statistical features, dynamic features, and fused features. The first hidden layer has 64 nodes, each node including the ReLU activation function and batch normalization. The second hidden layer has 32 nodes, each node including the ReLU activation function and batch normalization. The third hidden layer has 16 nodes, each node including the ReLU activation function. The output layer is a single node, including linear activation, used to output the roughness prediction bias. S5: Based on the statistical features, dynamic features, and fusion features, the roughness prediction distribution of the entire contour of the non-circular curve hole is obtained using the hybrid prediction model. The angle region corresponding to the preset threshold where the roughness in the roughness prediction distribution exceeds the average roughness of the entire contour is taken as the roughness anomaly region.

[0039] In one exemplary embodiment, the preset threshold is 20%-30%.

[0040] In one exemplary embodiment, the method for predicting the surface roughness of an inner circular irregular hole based on grinding force signals further includes: For each preset number of workpieces processed, one workpiece is randomly selected and its roughness is measured offline at a preset number of characteristic angle positions to obtain the measured value. The characteristic angle positions include curvature extreme points and curvature inflection points. The difference between the measured value and the predicted value of the hybrid prediction model is used as feedback. The weights of the neural network model layer of the hybrid prediction model are updated using incremental learning. Kalman filtering is used to correct the material correlation coefficient, curvature sensitivity coefficient and workpiece basic roughness of the physical model layer of the hybrid prediction model online, so as to obtain the updated hybrid prediction model. In one exemplary embodiment, the number of preset workpieces is 5-10; the number of preset angles is 3-5.

[0041] This embodiment provides a surface roughness prediction system for grinding internal irregular holes based on grinding force signals, including the following modules: The signal preprocessing module is used to: preprocess the triaxial grinding force signal in the non-circular curve hole grinding process to obtain the preprocessed triaxial grinding force signal, and perform angular domain equal-angle resampling on the preprocessed force signal according to the spindle rotation angle signal to convert the time domain triaxial grinding force signal into the angular domain triaxial grinding force signal. The geometric parameter calculation module is used to calculate the geometric parameters of non-circular curve holes at various angular positions, including the radius of curvature, equivalent grinding wheel diameter, and grinding contact arc length. The feature extraction module is used to: segment the triaxial grinding force signal of the angle domain according to the angle window of the preset width, and extract the statistical features, dynamic features and the fusion features of the statistical features and the geometric parameters of each angle window; The hybrid prediction model module is used to: construct a hybrid prediction model including a physical model layer and a neural network model layer. The physical model layer is used to obtain a prior physical prediction including a curvature-sensitive compensation term based on the theoretical relationship between grinding force and surface roughness. The neural network model layer is used to obtain the roughness prediction deviation based on the statistical features, dynamic features, and fusion features. The abnormal region identification module is used to: obtain the roughness prediction distribution of the entire contour of the non-circular curve hole based on the statistical features, dynamic features and fusion features using the hybrid prediction model, and identify the angle region corresponding to the roughness exceeding the preset threshold of the roughness of the entire contour in the roughness prediction distribution as the roughness abnormal region.

[0042] In some embodiments, the above-described method for predicting the surface roughness of grinding internal circular irregular holes based on grinding force signals can also be implemented in the following ways.

[0043] In this embodiment, the method for predicting the surface roughness of an inner circular irregular hole based on grinding force signals includes the following steps: (1) A force measuring instrument was used to collect the triaxial grinding force signal during the non-circular curve hole grinding process, including the normal force. Tangential force and axial force Synchronously acquire spindle rotation angle signals ; (2) The collected grinding force signal is preprocessed, including removing the trend term, low-pass filtering, and angle domain resampling of the force signal according to the spindle rotation angle signal, so as to convert the time domain force signal into the angle domain force signal. ; (3) Calculate the geometric parameters of the non-circular curved hole at each angular position, including the radius of curvature. Equivalent grinding wheel diameter Contact arc length with grinding ; (4) Window at the preset angle Force signals in the angular domain The process is segmented, and the statistical features, dynamic features, and fusion features of statistical features and geometric parameters of each window are extracted. (5) Constructing a hybrid prediction model, wherein the hybrid prediction model includes: The physical model layer establishes prior physical predictions based on the theoretical relationship between grinding force and surface roughness. It also includes curvature-sensitive compensation terms; The neural network model layer takes the features extracted in step (4) as input and outputs the roughness prediction bias. ; Hybrid prediction output: ; (6) Input the features of each angle window extracted in step (4) into the hybrid prediction model in step (5) in sequence to obtain the roughness prediction distribution of the full contour of the non-circular curve hole. , ; (7) Identify roughness anomaly regions based on predicted distribution: when When this occurs, the angular area is marked as an abnormal area and an alarm is output. The average roughness of the entire profile. This is a preset threshold.

[0044] Specifically, the non-circular curve hole includes any one of the following: elliptical hole, parabolic hole, hyperbolic hole, involute hole, and B-spline curve hole; the contour of the non-circular curve hole is defined by polar coordinate equations. It can be given in the form of a parametric equation.

[0045] Specifically, the low-pass filter cutoff frequency in step (2) is 500Hz; the number of sampling points for the angle domain equal-angle resampling is N=360 per revolution. 720 sampling points, with an angular resolution of 0.5°-1°.

[0046] Specifically, the geometric parameters in step (3) are calculated according to the following formula: Radius of curvature:

[0047] Equivalent grinding wheel diameter:

[0048] Grinding contact arc length: in The diameter of the grinding wheel. This refers to the grinding depth.

[0049] Specifically, the angular window width in step (4) The overlap rate between adjacent windows is 50%; the features extracted from each window include: Statistical characteristics: mean force in each direction , , Root mean square of forces in all directions , , , force ratio

[0050] Dynamic characteristics: coefficient of variation Peak frequency of force signal spectrum and peak amplitude

[0051] Fusion characteristic: unit normal force Grinding specific energy

[0052] Specifically, the physical model layer in step (5) adopts the following formula:

[0053] in: For material correlation coefficients, 0.15-0.25 is used for high-temperature alloys and 0.20-0.30 is used for titanium alloys; This is the curvature sensitivity coefficient, with a value ranging from 0.1 to 0.3. Used as the reference radius of curvature; This refers to the basic roughness of the workpiece.

[0054] Specifically, the neural network model layer in step (5) is a deep neural network, and its structure includes: Input layer: Dimensions 8-15, corresponding to the features extracted in step (4); First hidden layer: 64 nodes, activation function ReLU, batch normalization, Dropout=0.2; Second hidden layer: 32 nodes, activation function ReLU, batch normalization, Dropout=0.2; Third hidden layer: 16 nodes, activation function ReLU, Dropout=0.2; Output layer: Single node, linear activation, output roughness prediction bias .

[0055] Specifically, the above method also includes an online model update step: For every 5-10 workpieces processed, one workpiece is selected for offline roughness measurement at 3-5 characteristic angle positions, including curvature extrema and curvature inflection points. The difference between the measured value and the predicted value is used as feedback, and the weights of the neural network model layers are updated using an incremental learning method. Update strategy: Mini-batch gradient descent, batch size = 8-16, learning rate set to 10%-20% of the original model training rate; Physical model layer parameters , , Kalman filtering is used for online correction.

[0056] Specifically, the threshold in step (7) Select 20%-30%; when an abnormal area is identified, output the angle range of the abnormal area and the suggested process parameter compensation value.

[0057] In some embodiments, the above-described system for predicting the surface roughness of grinding internal circular irregular holes based on grinding force signals can also be implemented in the following ways.

[0058] In this embodiment, the surface roughness prediction system for grinding internal circular irregular holes based on grinding force signals includes: A triaxial piezoelectric force gauge, installed between the workpiece fixture and the machine tool table, has a force measurement range of 0-500N and is used to collect normal force. Tangential force and axial force ; The spindle encoder, coaxially mounted with the grinding machine spindle, is used to output spindle rotation angle signals. ; A multi-channel data acquisition card, connected to the force gauge and spindle encoder, with a sampling rate of not less than 2 kHz, is used to synchronously acquire force signals and rotation angle signals; The signal preprocessing module is used to perform detrending, low-pass filtering, and angle resampling in the angle domain on the force signal; The geometric parameter calculation module is used to calculate the radius of curvature, equivalent grinding wheel diameter, and grinding contact arc length at each angular position based on the non-circular curve hole profile equation. The feature extraction module is used to extract statistical features, dynamic features, and fused features of force signals according to the angle window. The hybrid prediction model module, which includes a physical model layer and a deep neural network layer, is used to output roughness prediction values ​​at various angular positions. The abnormal area identification and alarm module is used to identify areas with abnormal roughness and output process adjustment suggestions; The results display module is used to display the predicted distribution map of the full profile roughness.

[0059] In some embodiments, the above-described method for predicting the surface roughness of grinding internal circular irregular holes based on grinding force signals can also be implemented in the following ways.

[0060] like Figure 2 The diagram illustrates the complete process from signal acquisition to roughness prediction and anomaly identification, including key steps such as data acquisition, preprocessing, geometric parameter calculation, feature extraction, hybrid model prediction, full contour output, and model updating. In this embodiment, the method for predicting the surface roughness of an inner circular irregular hole based on grinding force signals includes the following steps: Step 1: Grinding force signal acquisition and angle domain mapping like Figure 3 The diagram illustrates the positional relationship between the force gauge mounted on the workpiece fixture and the machine tool table, the definition of the workpiece coordinate system, and the installation method of the spindle encoder. The force gauge is used to acquire triaxial grinding force signals, including the normal force, during the grinding process of non-circular curve holes. Tangential force and axial force The sampling rate should be no less than 2 kHz. The spindle rotation angle signal should be acquired synchronously. .

[0061] The acquired force signals are preprocessed, including: (1) Remove trend terms and eliminate zero-point drift; (2) Low-pass filtering with a cutoff frequency of 500 Hz to eliminate high-frequency noise interference; (3) Based on the spindle rotation angle signal The force signal is resampled at equal angles in the angle domain to convert the time-domain force signal F(t) into an angle-domain force signal. Each grinding cycle (0°-360°) collects N=360 data points. 720 sampling points ensure an angular resolution better than 0.5°-1°.

[0062] like Figure 4As shown, taking an elliptical aperture as an example, it demonstrates how the normal force signal Fn(t) acquired in the time domain is resampled into the angle domain signal Fn(θ) according to the principal axis rotation angle θ(t), and the force signal characteristics corresponding to the major axis endpoint and the minor axis endpoint are marked.

[0063] Step 2: Calculation of geometric parameters for non-circular curve holes For a given non-circular curve hole profile equation (Polar coordinates) Calculate the basic geometric parameters for each angular position: (1) Radius of curvature :

[0064] in and They are respectively right The first and second derivatives.

[0065] like Figure 5 As shown, taking an ellipse with a major semi-axis of a = 25 mm and a minor semi-axis of b = 15 mm as an example, the radius of curvature is illustrated. The variation pattern within the range of 0°-360° is marked, indicating the maximum curvature at the major axis endpoints (0°, 180°) and the minimum curvature at the minor axis endpoints (90°, 270°). (2) Equivalent grinding wheel diameter :

[0066] in The diameter is the grinding wheel diameter.

[0067] (3) Grinding contact arc length :

[0068] in This refers to the grinding depth.

[0069] Step 3: Angular Window Feature Extraction like Figure 6 As shown, the mean normal force F is illustrated. _n, mean tangential force F The curves showing the variation of force _t and force ratio R_nt with angle θ demonstrate the correspondence between the force signal and the non-circular curve hole profile. Window at preset angle Force signals in the angular domain Segmentation processing, window width The overlap rate between adjacent windows is 50%. The following features are extracted for each window: (1) Statistical characteristics: Mean normal force:

[0070] Mean tangential force:

[0071] Mean axial force:

[0072] Root mean square of forces in all directions: , ,

[0073] Force ratio:

[0074] (2) Dynamic characteristics: Coefficient of variation:

[0075] Force signal spectrum peak frequency and peak amplitude

[0076] (3) Fusion features (combination of statistical features and geometric parameters): Unit normal force:

[0077] Grinding specific energy:

[0078] Step 4: Construct a hybrid prediction model like Figure 7 As shown, the parallel architecture of the physical model layer (including curvature-sensitive compensation terms) and the deep neural network layer (three hidden layers) and their fusion output method are illustrated; a hybrid prediction model integrating physical mechanisms and data-driven approaches is established. (4.1) Physical Model Layer Based on the theoretical relationship between grinding force and surface roughness, a priori physical prediction is established:

[0079] in: For material correlation coefficients, 0.15-0.25 is used for high-temperature alloys and 0.20-0.30 is used for titanium alloys; This is the curvature sensitivity coefficient, with a value ranging from 0.1 to 0.3, used to compensate for the additional effects of curvature changes on roughness; Use the reference radius of curvature (such as the major semi-axis or average radius of curvature of an elliptical hole). This refers to the basic surface roughness of the workpiece.

[0080] (4.2) Neural Network Model Layer A feature-roughness deviation mapping is established using a deep neural network: Input layer: Feature vectors extracted in step 3, dimension D=8 15D=8 15; First hidden layer: 64 nodes, activation function ReLU, batch normalization, dropout=0.2; Second hidden layer: 32 nodes, activation function ReLU, batch normalization, Dropout=0.2; Third hidden layer: 16 nodes, activation function ReLU, Dropout=0.2; Output layer: Single node, linear activation, output roughness prediction bias ; (4.3) Hybrid Prediction Output

[0081] Step 5: Full contour roughness distribution prediction and anomaly region identification like Figure 8 As shown, the method of the present invention predicts... The curves were compared with the offline measured values, and the error at each angle position was marked. The features of each angle window extracted in step 3 are sequentially input into the hybrid prediction model in step 4 to obtain the roughness prediction distribution of the entire contour of the non-circular curve hole. , .

[0082] Analyze the prediction results: Calculate the mean roughness of the entire profile: ; Identify roughness anomaly areas: The area, in which The threshold (taken as 20%-30%). Output the angle range of the abnormal area and the suggested process parameter compensation values; like Figure 9 As shown, the predicted roughness distribution of the entire profile of the parabolic hole is displayed in the form of a polar coordinate heatmap, with different colors representing different roughness levels. Step 6: Online Model Update like Figure 10 As shown, the collaborative update mechanism of incremental learning and Kalman filtering is illustrated, including triggering conditions, feature angle measurement, model parameter update, and other steps. For every 5-10 workpieces processed, one workpiece is selected for offline roughness measurement at 3-5 characteristic angle positions (such as curvature extreme points and inflection points); The difference between the measured value and the predicted value is used as feedback, and the weights of the neural network model layers are updated using an incremental learning method. Update strategy: Mini-batch gradient descent, batch size = 8-16, learning rate set to 10%-20% of the original model training rate; Physical model layer parameters , , Kalman filtering is used for online correction.

[0083] In some embodiments, the above-described method for predicting the surface roughness of grinding internal circular irregular holes based on grinding force signals can also be implemented in the following ways.

[0084] This embodiment focuses on predicting the surface roughness of an elliptical hole (GH4169) during grinding; the experimental conditions are as follows: Workpiece material: GH4169 high temperature alloy (quenched and tempered, hardness approximately 40 HRC); Hole type: elliptical hole, major semi-axis = 25mm, minor semi-axis b = 15mm, hole depth 30mm; Polar equation of elliptical profile:

[0085] Grinding machine: Five-axis linkage CNC internal grinding machine (maximum spindle speed 60,000 rpm); Force measuring instrument: Triaxial piezoelectric force measuring instrument (range 0-500 N), installed between the workpiece fixture and the machine tool table; Grinding wheel: CBN internal grinding wheel (diameter) =20mm, particle size B151, concentration 150%). Process parameters: Grinding wheel linear velocity =45m / s; Workpiece rotation speed =500rpm (constant angular velocity); Grinding depth ==0.01mm / pass; Coolant: Water-based emulsion, pressure 10 bar; Signal acquisition: Sampling rate 5kHz, synchronously acquiring spindle encoder rotation angle signal; Force signal acquisition and angle domain conversion: The triaxial force signal is continuously acquired throughout the entire grinding cycle (approximately 120 ms, corresponding to 1 revolution). After the original time-domain signal is processed by detrending and low-pass filtering (cutoff frequency 500 Hz), it is resampled in the angle domain according to the rotation angle signal output by the spindle encoder, with 720 points acquired per revolution and an angle resolution of 0.5°.

[0086] The measured force signal values ​​at the characteristic angle are as follows: Geometric parameter calculation: Calculate the geometric parameters of each angle using the polar equation of an ellipse. =0° (end of major axis) and For example, 90° (minor axis endpoint): =0°: =25.0mm, =2 / (1 / 20+1 / 25)=22.2mm, =0.01×22.2=0.471mm; =90°: =15.0mm, =2 / (1 / 20+1 / 15)=17.1mm, =0.01×17.1=0.414mm; 4. Angle Window Feature Extraction Set window width =10°, sliding step size 5°, a total of 72 windows were obtained. For example, a 0° window (-5° to +5°): Statistical characteristics: ; Dynamic characteristics: =235Hz, =3.2N; Fusion characteristics: =121.5 / 0.471=258N / mm; Material removal rate: MRR= =0.01×(500×2π×25 / 60)×30=392.7mm³ / s; Grinding specific energy: u = 48.6 × 45 / 392.7 = 5.57 J / mm³; Model prediction: (1) Physical model prediction (taking θ=0° window as an example) Pick =0.18, =0.2, =20mm (average radius of curvature of the ellipse). =0.10μm:

[0087] After inverse normalization (training data range 0.2-0.5μm): =0.38μm; (2) Neural Network Bias Prediction The neural network input feature dimension is 9-dimensional: [ ]; Network output: = 0.01μm; (3) Hybrid prediction output =0.38 0.01 = 0.37 μm; Full contour prediction results and validation: The roughness prediction curve for the entire profile was obtained by sequentially calculating the roughness of 72 angular windows. Verification was performed by measuring one point every 15° along the elliptical hole profile (24 points in total), as shown in Table 1. Table 1: Full contour prediction results and validation table

[0088] Overall accuracy: Mean absolute error (MAE) = 0.0083 μm; Mean relative error (MRE) = 2.7%; The maximum relative error is 3.6% (occurring in the region at the minor axis endpoint where the curvature changes most drastically). The coefficient of determination, R², is 0.967. Abnormal region identification: Set threshold =25%, average roughness of the entire profile =0.34μm, threshold upper limit = 0.34 × 1.25 = 0.425μm. All predicted values ​​are below the threshold upper limit, and there are no outlier regions.

[0089] Online model update and validation: After machining the 10th workpiece, offline measurements were performed at the major axis endpoint (0°) and the minor axis endpoint (90°). The differences between the measured and predicted values ​​were +0.01μm and -0.02μm, respectively. Incremental learning was used to update the neural network weights (learning rate 0.001, batch size=8). After the update, the average prediction error for the 11th-15th workpieces decreased from 0.012μm to 0.008μm.

[0090] like Figure 11 As shown, the mean relative error (MRE) and coefficient of determination (R2) of the method of the present invention are compared with those of the pure data-driven neural network method and the traditional empirical formula method under different working conditions.

[0091] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A method for predicting the surface roughness of an inner circular irregular hole based on grinding force signals, characterized in that, Includes the following steps: S1: The triaxial grinding force signal in the non-circular curve hole grinding process is preprocessed to obtain the preprocessed triaxial grinding force signal. The preprocessed force signal is then resampled in the angle domain according to the spindle rotation angle signal to convert the time domain triaxial grinding force signal into the angle domain triaxial grinding force signal. S2: Calculate the geometric parameters of the non-circular curve hole at each angular position, including the radius of curvature, equivalent grinding wheel diameter, and grinding contact arc length; S3: The triaxial grinding force signal in the angle domain is segmented according to the angle window of the preset width, and the statistical features, dynamic features, and fusion features of the statistical features and the geometric parameters of each angle window are extracted. S4: Construct a hybrid prediction model including a physical model layer and a neural network model layer. The physical model layer is used to obtain a prior physical prediction including a curvature-sensitive compensation term based on the theoretical relationship between grinding force and surface roughness. The neural network model layer is used to obtain the roughness prediction deviation based on the statistical features, dynamic features and fusion features. S5: Based on the statistical features, dynamic features, and fusion features, the roughness prediction distribution of the entire contour of the non-circular curve hole is obtained using the hybrid prediction model. The angle region corresponding to the preset threshold where the roughness in the roughness prediction distribution exceeds the average roughness of the entire contour is taken as the roughness anomaly region.

2. The method according to claim 1, characterized in that, The non-circular curve hole includes at least one of elliptical hole, parabolic hole, hyperbolic hole, involute hole and B-spline curve hole; the triaxial grinding force signal includes normal force, tangential force and axial force.

3. The method according to claim 1, characterized in that, The preprocessing includes removing trend terms, eliminating zero-point drift, and low-pass filtering.

4. The method according to claim 1, characterized in that, The formula for calculating the geometric parameters is as follows: in, The radius of curvature; The equivalent grinding wheel diameter; This is the grinding contact arc length; For polar angle In terms of direction, the distance from the pole to the contour point of the inner wall of the hole; and They are respectively First and second derivatives with respect to θ; The diameter of the grinding wheel; This refers to the grinding depth.

5. The method according to claim 1, characterized in that, The statistical features include mean normal force, mean tangential force, mean axial force, root mean square of normal force, root mean square of tangential force, root mean square of axial force, and force ratio; the dynamic features include coefficient of variation, peak frequency and peak amplitude of the three-axis grinding force signal; the fusion features of the statistical features and the geometric parameters include unit normal force and grinding specific energy.

6. The method according to claim 1, characterized in that, The expression for the hybrid prediction output of the hybrid prediction model is: in, For the hybrid prediction model for the first Mixed prediction output from multiple angle windows; For the first Prior physics predictions for each angle window; For the first Roughness prediction bias for each angle window.

7. The method according to claim 1, characterized in that, The calculation formula for the prior physical prediction is as follows: in: For the first Prior physics predictions for each angle window; Material correlation coefficient; For the first The average tangential force of each angle window; The linear velocity of the grinding wheel; For the first Workpiece rotation speed at each angle window; This refers to the grinding depth. The curvature sensitivity coefficient; For the first The radius of curvature of each angle window; Used as the reference radius of curvature; This refers to the basic roughness of the workpiece.

8. The method according to claim 1, characterized in that, The neural network model layer is a deep neural network, and the structured deep neural network includes an input layer, a first hidden layer, a second hidden layer, a third hidden layer, and an output layer; the input layer has a dimension of 8-15 and is used to receive the statistical features, dynamic features, and fused features; The first hidden layer has 64 nodes, each node including the ReLU activation function and batch normalization; the second hidden layer has 32 nodes, each node including the ReLU activation function and batch normalization; the third hidden layer has 16 nodes, each node including the ReLU activation function; the output layer is a single node, including linear activation, used to output the roughness prediction bias.

9. The method according to claim 1, characterized in that, Also includes: For each preset number of workpieces processed, one workpiece is randomly selected and its roughness is measured offline at a preset number of characteristic angle positions to obtain the measured value. The characteristic angle positions include curvature extreme points and curvature inflection points. The difference between the measured value and the predicted value of the hybrid prediction model is used as feedback. The weights of the neural network model layer of the hybrid prediction model are updated using incremental learning. Kalman filtering is used to correct the material correlation coefficient, curvature sensitivity coefficient and workpiece basic roughness of the physical model layer of the hybrid prediction model online, so as to obtain the updated hybrid prediction model.

10. A surface roughness prediction system for grinding internal circular irregular holes based on grinding force signals, characterized in that, The system includes the following modules: The signal preprocessing module is used to: preprocess the triaxial grinding force signal in the non-circular curve hole grinding process to obtain the preprocessed triaxial grinding force signal, and perform angular domain equal-angle resampling on the preprocessed force signal according to the spindle rotation angle signal to convert the time domain triaxial grinding force signal into the angular domain triaxial grinding force signal. The geometric parameter calculation module is used to calculate the geometric parameters of non-circular curve holes at various angular positions, including the radius of curvature, equivalent grinding wheel diameter, and grinding contact arc length. The feature extraction module is used to: segment the triaxial grinding force signal of the angle domain according to the angle window of the preset width, and extract the statistical features, dynamic features and the fusion features of the statistical features and the geometric parameters of each angle window; The hybrid prediction model module is used to: construct a hybrid prediction model including a physical model layer and a neural network model layer. The physical model layer is used to obtain a prior physical prediction including a curvature-sensitive compensation term based on the theoretical relationship between grinding force and surface roughness. The neural network model layer is used to obtain the roughness prediction deviation based on the statistical features, dynamic features, and fusion features. The abnormal region identification module is used to: obtain the roughness prediction distribution of the entire contour of the non-circular curve hole based on the statistical features, dynamic features and fusion features using the hybrid prediction model, and identify the angle region corresponding to the roughness exceeding the preset threshold of the roughness of the entire contour in the roughness prediction distribution as the roughness abnormal region.

Images