Method for monitoring damage of hard and brittle material in ultra-precision turning based on multi-sensor fusion

Abstract

The application discloses a hard and brittle material ultra-precision turning damage monitoring method based on multi-sensor fusion, relates to the precision / ultra-precision machining state monitoring technical field, and builds a multi-sensor signal acquisition device on an ultra-precision turning machine tool to capture different types of sensor signals generated in a turning process in real time; the acquired sensor signals are subjected to a signal pretreatment operation of minimum entropy deconvolution filtering; and 95 time domain and frequency domain statistical features reflecting turning damage are extracted from the signals subjected to the minimum entropy deconvolution filtering. The application significantly improves the precision and reliability of monitoring by integrating data from different sensors, captures acoustic features, cutting force and mechanical vibration information in the machining process, and fuses the information, so that more comprehensive and accurate information is obtained, the wear degree of a tool and the damage condition of a workpiece are more accurately judged, false positives and false negatives are reduced, and the reliability of monitoring is improved.

Description

Technical Field

[0001] This invention relates to the field of precision / ultra-precision machining condition monitoring technology, specifically to a method for monitoring damage during ultra-precision turning of hard and brittle materials based on multi-sensor fusion. Background Technology

[0002] Hard and brittle materials such as single-crystal silicon, single-crystal germanium, glass, silicon carbide, and sapphire have shown great application potential in high-tech fields such as aerospace, electronic engineering, and optics due to their excellent physical and chemical properties. As modern science and technology continue to increase the requirements for material processing precision, the limitations of traditional grinding technology have become increasingly apparent, specifically manifested in problems such as low processing efficiency, high cost, and susceptibility to thermal damage during processing. In contrast, ultra-precision turning technology, through the synergistic effect of diamond tools, high-precision CNC machine tools, and precision measurement systems, has achieved micron- to nanometer-level processing precision while maintaining high processing efficiency. Currently, ultra-precision turning technology has been widely and deeply applied in the field of hard and brittle material processing.

[0003] In the electronics industry, this technology is used to manufacture high-performance alumina ceramic capacitors and insulators, significantly improving the performance and reliability of electronic devices. In the field of infrared optics, ultra-precision turning technology is used to process optical components made of materials such as sapphire, fused silica, and single-crystal silicon, including lenses, mirrors, and windows. These components play a crucial role in infrared detection, imaging, and communication. However, the inherent high hardness, brittleness, and low thermal conductivity of hard and brittle materials make them highly susceptible to brittle damage such as cracks and spalling during ultra-precision turning, which affects the integrity of the machined surface and makes it difficult to meet the stringent performance requirements of high-end applications. Currently, suppressing / eliminating brittle damage caused by the processing of hard and brittle materials has become a core requirement in this field.

[0004] To further improve machining quality, increase production efficiency, and reduce manufacturing costs, the application of intelligent monitoring systems in the field of ultra-precision turning is gradually gaining attention. However, current signal monitoring for ultra-precision turning mainly focuses on tool wear, chatter, and workpiece roughness, and there is a lack of effective real-time signal monitoring methods to predict damage to the material machining surface. In addition, current signal monitoring processes often rely on a single sensor, and the information obtained is relatively one-sided and contains large errors. Given the increasing demand for ultra-precision non-destructive machining of hard and brittle materials, proposing a machining damage in-situ monitoring method based on multi-sensor signals has important practical significance. Summary of the Invention

[0005] The purpose of this invention is to provide a method for monitoring damage during ultra-precision turning of hard and brittle materials based on multi-sensor fusion, so as to solve the problems mentioned in the background art.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0007] A method for monitoring damage during ultra-precision turning of hard and brittle materials based on multi-sensor fusion includes the following steps:

[0008] Step 1: Build a multi-sensor signal acquisition device on the ultra-precision turning machine tool, including acoustic emission sensors, force sensors and vibration sensors. The sensors are all set at the tool end near the material processing area to capture different types of sensor signals generated during the turning process in real time.

[0009] Step 2: Perform minimum entropy deconvolution (MED) filtering on the acquired sensor signals for signal preprocessing to enhance signal recognizability, reduce adverse effects, and restore the impulse components in the filtered signal to the greatest extent possible.

[0010] Step 3: Extract features from the signal after minimum entropy deconvolution filtering. Extract a total of 95 time-domain and frequency-domain statistical features reflecting turning damage, including time-domain statistical indicators such as peak value, mean, root mean square value, standard deviation, kurtosis, and skewness, as well as frequency-domain statistical indicators such as spectral centroid and mean square frequency.

[0011] Step 4: Use kernel principal component analysis to perform feature dimensionality reduction and fusion. Through min-max normalization and radial basis function kernel, map the original data to a high-dimensional feature space, calculate the kernel matrix of the feature matrix, perform feature decomposition, select the top few principal components whose cumulative contribution rate reaches a certain threshold, project the original data onto the selected principal components, and then obtain the dimensionality-reduced data.

[0012] Step 5: The least squares support vector machine method optimized by genetic algorithm is used for pattern recognition to identify the material removal patterns of hard and brittle materials in the ultra-precision turning process and to determine the state of machining damage.

[0013] A further improvement to the technical solution of the present invention is that: in step 1, the process of capturing the sensor signal during the turning process is as follows:

[0014] Step 101: Install an acoustic emission sensor, a force sensor, and a vibration sensor on the cutting edge of the ultra-precision turning machine tool near the material processing area to capture acoustic characteristics, cutting force, and mechanical vibration information during the processing, respectively. Ensure that the sensors are fixed and stable to avoid displacement or vibration during the turning process, which would affect the accuracy of the signal. Install a signal preprocessing device, including a current amplifier and a filter. The current amplifier is installed to enhance the weak signal output by the sensor, and the filter is installed to reduce noise and interference and improve signal quality.

[0015] Step 102: Connect the signal line output by the sensor to the data acquisition card of the data acquisition module to ensure that the data acquisition card is correctly connected and compatible with the host computer. Configure the sampling rate and resolution of the data acquisition card to ensure that the digital signal can accurately reflect the physical quantity measured by the sensor. Set the input channel of the data acquisition card to correspond to the signal input of each sensor.

[0016] Step 103: Start the data acquisition card to continuously acquire the output signals of each sensor during the turning process. After the sensor signals are processed by the current amplifier and filter, they are transmitted to the data acquisition card of the data acquisition module for analog-to-digital (A / D) conversion.

[0017] Step 104: Store the acquired digital signals on the hard drive of the host computer to ensure the stability and security of data transmission and avoid data loss or damage for subsequent processing and analysis. Install the corresponding signal processing software on the host computer to receive, store and preliminarily process the acquired digital signals. At the same time, perform a preliminary check on the data from the perspectives of signal integrity, synchronization and timestamp verification.

[0018] A further improvement to the technical solution of the present invention is that, in step 2, the signal preprocessing process is as follows:

[0019] Step 201: Load the collected force signal, vibration signal and acoustic emission signal data from the host computer's hard disk, and perform preprocessing operations on the loaded signal data to remove DC offset, filter out low-frequency drift and high-frequency noise.

[0020] Step 202: Set the length L of the inverse filter according to the characteristics and requirements of the signal, initialize the parameters of the inverse filter with random values, and set the objective function as minimum entropy. Then calculate the impact signal generated during the contact process between the tool and the workpiece. The longer the length of the inverse filter, the more signal features can be captured, but the computational complexity will also increase. Minimum entropy means that the entropy of the output signal is the smallest, that is, the information in the signal is the most concentrated.

[0021] Step 203: Use normalized kurtosis as the objective function to optimize the inverse filter. Through iterative calculation, adjust the parameters of the inverse filter to maximize the kurtosis, thereby achieving the minimum entropy condition. Then apply the optimized inverse filter to the original signal to determine the optimal inverse filter matrix.

[0022] Step 204: Perform time-domain and frequency-domain analysis on the filtered signal. Observe the changes in the signal waveform and impulse components through time-domain analysis, and observe the changes in the signal spectrum and frequency components through frequency-domain analysis. Compare the time-domain waveforms and frequency-domain spectra of the signals before and after filtering to verify whether the impulse components are significantly enhanced and whether the signal-to-noise ratio is improved. Compare the time-domain waveforms and frequency-domain spectra of the original signal and the filtered signal to visually demonstrate the filtering effect. This helps to confirm whether the minimum entropy deconvolution filtering has successfully extracted the impulse components in the signal and weakened the adverse effects.

[0023] A further improvement of the technical solution of the present invention is that the calculation formula for the impact signal generated during the contact process between the tool and the workpiece is expressed as follows:

[0024]

[0025] Where y(n) is the impact signal generated during the contact process between the tool and the workpiece, f(n) is the inverse filter, x(n) is the original acquired signal, L is the length of the inverse filter, * is the convolution operation, f in f(l) refers to the convolution / inverse filter set in the minimum entropy deconvolution filtering method, which is equivalent to a convolution kernel, n is the nth sample in the signal time series, and l is the lth coefficient in the convolution kernel;

[0026] Since a larger kurtosis value (the fourth-order center moment of the normalized data) results in a smaller entropy value, the normalized kurtosis is chosen as the objective function in minimum entropy deconvolution filtering (MED), and its expression is:

[0027]

[0028] When the formula is When the maximum value is reached, the extreme value condition is satisfied. Then it can be determined that f(l) is the optimal inverse filter, and the following equation can be obtained by combining formula (1):

[0029]

[0030] in, For normalized kurtosis, y(i) represents the i-th sample in the time series of the output impulse signal, x, y, and f functions are the original acquired signal, the output impulse signal, and the convolution / inverse filter (convolution kernel) set in the minimum entropy deconvolution filtering (MED) method, respectively. n and l represent the n-th or l-th sample in the signal time series, p represents the p-th coefficient in the convolution kernel, and i, n, l, p represent the i-th, n, l, and p-th values ​​in the x, y, and f functions, respectively.

[0031] Formula (3) can be written in matrix form:

[0032] f = A -1×b; (4)

[0033] The optimal inverse filter matrix is ​​finally determined by iterative calculation using formula (4).

[0034] A further improvement to the technical solution of this invention lies in the following: In step 3, the extraction process of the time-domain and frequency-domain statistical features of turning damage is as follows:

[0035] Step 301: Import the signal data processed by minimum entropy deconvolution filtering into the signal analysis software, extract the turning damage feature values ​​of the different sensor filtered signals, which are time domain and frequency domain statistical features, to provide input variables for subsequent pattern recognition or machine learning models.

[0036] Step 302: Perform time-domain analysis on the filtered signals from different sensors and extract 11 time-domain statistical indicators from the time-domain signals, including peak value, mean, root mean square value, standard deviation, kurtosis, skewness, peak-to-peak value, peak factor, impulse factor, waveform factor, and margin factor.

[0037] Step 303: Perform a Fast Fourier Transform on the time-domain signal to obtain the signal spectrum information, and extract eight frequency domain statistical indicators, including the spectral centroid, mean square frequency, root mean square frequency, frequency variance, spectral peak value, power spectral entropy, spectral kurtosis, and spectral skewness.

[0038] Step 304: Integrate the time-domain and frequency-domain features extracted from each sensor to form a feature vector, and extract a total of 95 statistical features, including force signals in the X, Y, and Z directions.

[0039] A further improvement to the technical solution of this invention lies in the following: In step 4, the process of obtaining the dimensionality-reduced data is as follows:

[0040] Step 401: The 95 extracted statistical features are organized into a feature matrix, with each row representing a sample and each column representing a feature. The feature matrix is ​​then subjected to Min-Max standardization to scale the value of each feature to the range of [0,1], eliminating the influence of different dimensions between features and ensuring that the data are of the same magnitude. Kernel function calculation is then performed.

[0041] Step 402: Select the radial basis function (RBF) as the kernel function. The RBF kernel is also called the Gaussian kernel and is suitable for mapping nonlinear data. Using the selected radial basis function (RBF) kernel function, calculate the kernel function value between each sample point in the original dataset to form a kernel matrix. Each element of the kernel matrix is ​​the kernel function value between two sample points.

[0042] Step 403: Subtract the mean of the corresponding row and column from each element in the kernel matrix to achieve kernel matrix centering, ensure the correct mapping of data in high-dimensional space, eliminate the offset in the data, and make the subsequent feature decomposition more accurate.

[0043] Step 404: Perform eigenvalue decomposition on the centered kernel matrix to obtain eigenvalues ​​and corresponding eigenvectors. The eigenvalues ​​represent the amount of information contained in each principal component, and the eigenvectors represent the direction of the principal components. Based on the magnitude of the eigenvalues, select several (first few) principal components whose cumulative contribution rate reaches a certain threshold (such as 95% or 99%). These principal components can retain the information of the original data to the greatest extent.

[0044] Step 405: Project the original data onto the selected principal components. By multiplying the kernel matrix with the eigenvectors of the selected principal components, the dimensionality-reduced data is obtained. The dimensionality-reduced data not only retains the main variation information of the original data, but also greatly reduces the dimensionality, thereby effectively reducing redundant information, improving the computational efficiency of subsequent models, and outputting the dimensionality-reduced data. The dimensionality-reduced eigenvectors are the final fused feature set, which will be used as the input of subsequent machine learning models.

[0045] A further improvement to the technical solution of this invention lies in that: the calculation formula for the radial basis function (RBF) kernel function is as follows:

[0046]

[0047] Among them, K jk z is the inner product of sample j and sample k in the high-order feature space. j and z k It is the vector representation of two samples in the original feature space, and σ is the kernel width parameter.

[0048] A further improvement to the technical solution of the present invention is that: in step 5, the process for determining the state of processing damage is as follows:

[0049] Step 501: The reduced feature vector is used as input to match the material removal mode and damage state label corresponding to the hard and brittle material to obtain the reduced feature set, which is used to train and test the LSSVM model. The feature set is divided into training set and test set, with a ratio of 70% training set and 30% test set.

[0050] Step 502: Construct a least squares support vector machine (LSSVM) classification model, select the radial basis function (RBF) kernel function, and define the loss function;

[0051] Step 503: Use a genetic algorithm (GA) to optimize the hyperparameters (such as regularization parameters and kernel width) of the least squares support vector machine (LSSVM) classification model. The genetic algorithm generates multiple candidate solutions through selection, crossover and mutation operations, and evaluates the fitness of each solution. By iterating the genetic algorithm optimization process, the optimal combination of hyperparameters is found, thereby improving the classification performance of the least squares support vector machine (LSSVM) classification model.

[0052] Step 504: After determining the optimal hyperparameters, train the least squares support vector machine (LSSVM) classification model using the training set data. During training, the model learns how to map different input features to the corresponding output categories (i.e., plastic removal or brittle removal). After training, evaluate the model using the test set. Measure the model performance by calculating metrics such as classification accuracy, recall, precision, and F1-score. At the same time, the confusion matrix can be used to visualize the model's classification results.

[0053] Step 505: Input the sample data to be identified into the trained least squares support vector machine (LSSVM) classification model to obtain the identification result;

[0054] Step 506: Set the expected value of the feature value under normal turning conditions, combine the feature value of the statistical feature in the feature set after dimensionality reduction, calculate the turning damage identification coefficient, and then analyze the damage state of hard and brittle materials.

[0055] Step 507: Based on the identification results and damage state analysis results, determine the material removal mode of the hard and brittle material in the ultra-precision turning process, and determine the damage state in the processing process based on the identified material removal mode and the processing parameters and process conditions.

[0056] A further improvement to the technical solution of the present invention is that the calculation expression for the turning damage identification coefficient is:

[0057]

[0058] Where D is the turning damage identification coefficient, representing the damage state, g t Let t be the eigenvalue, which is the statistical feature of the feature set after dimensionality reduction. M is the total number of features. B is the expected value of the eigenvalue under normal turning conditions, which is determined through experiments or historical data analysis. P is the limit value, which represents the maximum allowable deviation between the eigenvalue and the benchmark value, and is used to adjust the sensitivity of the exponential function. The value of D ranges from 0 to 1, and the closer the value is to 1, the more severe the damage.

[0059] Due to the adoption of the above technical solution, the technical progress achieved by this invention compared to the prior art is as follows:

[0060] 1. This invention provides a method for monitoring damage in ultra-precision turning of hard and brittle materials based on multi-sensor fusion. By integrating data from different sensors, the accuracy and reliability of monitoring are significantly improved. By capturing acoustic features, cutting forces and mechanical vibration information during the machining process and fusing them, more comprehensive and accurate information can be obtained, thereby more accurately judging the wear degree of the tool and the damage condition of the workpiece. This helps to reduce false alarms and missed alarms and improve the reliability of monitoring.

[0061] 2. This invention provides a method for monitoring damage during ultra-precision turning of hard and brittle materials based on multi-sensor fusion. By monitoring tool wear and workpiece damage in real time, potential problems can be detected and addressed in a timely manner, avoiding workpiece scrap and increased processing costs. At the same time, by optimizing processing parameters, processing efficiency can be improved, processing time and energy consumption can be reduced. In addition, multi-sensor fusion technology can realize autonomous control of the processing process, reduce manual intervention and operating costs, and can significantly reduce processing costs and improve production efficiency. Attached Figure Description

[0062] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0063] Figure 1 This is a schematic diagram of the multi-sensor signal processing flow of the present invention;

[0064] Figure 2 This is a schematic diagram of the multi-sensor signal acquisition device for ultra-precision turning machine tools according to the present invention;

[0065] Figure 3 This is a schematic diagram of the material removal mode during the turning process of hard and brittle materials according to the present invention;

[0066] Figure 4 This is a comparison diagram of the original force signal and the force signal after MED filtering according to the present invention. Detailed Implementation

[0067] 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. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. 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.

[0068] Example 1, such as Figures 1 to 4As shown, this invention provides a method for monitoring damage during ultra-precision turning of hard and brittle materials based on multi-sensor fusion, comprising the following steps:

[0069] Step 1: Build a multi-sensor signal acquisition device on the ultra-precision turning machine tool, including acoustic emission sensors, force sensors, and vibration sensors. All sensors are positioned near the material processing area at the tool tip to capture different types of sensor signals generated during the turning process in real time. The output signals of each sensor are continuously acquired during the turning process. The acquired signals are transmitted to a data acquisition card for A / D conversion via pre-processing devices such as current amplifiers and filters. Digital signals are collected and input to a host computer for further processing. Acoustic emission sensors, force sensors, and vibration sensors are installed near the material processing area at the tool tip of the ultra-precision turning machine tool to capture acoustic characteristics, cutting force, and mechanical vibration information during the machining process, respectively. Ensure the sensors are fixed and stable to avoid displacement or vibration during the turning process, which could affect signal accuracy. A signal pre-processing device, including current amplifiers and filters, is also installed. The current amplifiers enhance the weak signals output by the sensors, while the filters reduce noise and interference, improving signal quality.

[0070] Furthermore, by using force sensors to monitor the cutting force applied by the tool to the workpiece in real time during the cutting process, the changes in the force can reflect the interaction between the tool and the workpiece. When brittle damage occurs on the workpiece surface, the cutting force will exhibit abnormal fluctuations. The threshold for abnormal fluctuations depends on the machining material, tool material, and machining parameters. Generally, multiple experiments are conducted on the cutting force under specific materials and cutting conditions to analyze the fluctuation characteristics of normal fluctuations and those occurring when brittle damage occurs, thereby establishing the threshold for abnormal fluctuations. According to previous literature, the difference between the toughness and brittleness of hard and brittle materials affects the amplitude of cutting force fluctuations. At the same time, changes in parameters such as cutting speed, feed rate, and depth of cut during the machining process also affect the amplitude of cutting force fluctuations. In addition, the wear degree of the tool itself also affects the amplitude of cutting force fluctuations. Under normal circumstances, the variation of cutting force during normal cutting is ±5% to ±15%. Exceeding this range indicates excessive tool wear, workpiece damage, or machine tool problems. Vibration sensors monitor the vibration state of the machine tool and tool during the cutting process. The threshold of mechanical vibration is related to the specific material being machined, the tool material, and the working conditions. The acceleration (m / s²) in the vibration sensor... 2The acceleration value (or g value) is a commonly used measurement index, representing the vibration amplitude during the cutting process. Under normal machining conditions, the acceleration value is usually kept within a certain range (such as 0.1 to 0.5g). If the peak acceleration detected by the vibration sensor exceeds about 1.0 to 1.5g, it means that the instability of the cutting process has increased and brittle damage may occur. However, the threshold varies depending on the material, tool, machining speed and process conditions, and needs to be adjusted according to the specific application and experimental results. Furthermore, changes in signal frequency components can effectively identify the characteristics of brittle damage. Brittle damage is often accompanied by an increase in high-frequency components (e.g., exceeding 5kHz). Through spectral analysis, if a significant increase in the amplitude of certain high-frequency components is found, it can be determined that the instability of the cutting process has increased, and damage has occurred on the machined surface. By capturing stress wave signals generated inside the workpiece during the cutting process using an acoustic emission sensor, when brittle damage occurs on the workpiece surface, the internal crystal structure of the material changes, thereby generating stress waves. The acoustic emission sensor can detect this signal. The signal range is determined according to the actual material and working conditions. The degree of brittle damage is determined by changes in the signal. For low-brittle damage, the internal crystal lattice damage is small, limited to the surface or micro-cracks, and does not significantly reduce the material's load-bearing capacity. In terms of processing quality, materials with low brittleness damage can generally be used in less demanding applications, such as structural components or parts in low-load environments. However, they cannot be used in high-end applications, such as infrared optical elements or single-crystal silicon mirrors, because such applications require very good surface quality (mirror-level). Even slight surface damage can seriously affect the surface quality. In contrast, medium to high brittleness damage indicates that the processing damage has extended to a deeper level, forming large cracks or even material spalling. The strength, toughness, and wear resistance of the material are significantly reduced, making it unsuitable for applications in environments with precision or high strength requirements, such as high-load load-bearing components or precision instrument parts. For materials that have reached high brittleness damage, they have usually lost their original functionality, and their applicability is severely limited, or even unusable.

[0071] Furthermore, the signal lines output by the sensors are connected to the data acquisition card to ensure proper connection and compatibility between the data acquisition card and the host computer. The sampling rate and resolution of the data acquisition card are configured to ensure that the digital signals accurately reflect the physical quantities measured by the sensors. The input channels of the data acquisition card are set to correspond to the signal inputs of each sensor. The data acquisition card is then started to continuously acquire the output signals of each sensor during the turning process. After the sensor signals are processed by the current amplifier and filter, they are transmitted to the data acquisition card of the data acquisition module for analog-to-digital (A / D) conversion. The acquired digital signals are stored on the hard drive of the host computer to ensure the stability and security of data transmission and to avoid data loss or damage for subsequent processing and analysis. Appropriate signal processing software is installed on the host computer to receive, store, and preliminarily process the acquired digital signals. At the same time, the data is preliminarily checked from the perspectives of signal integrity, synchronization, and timestamp verification.

[0072] Step 2 involves performing minimum entropy deconvolution (MED) filtering on the acquired sensor signals (force signals, vibration signals, and acoustic emission signals) to enhance signal identifiability, reduce adverse effects, and restore the impact components in the filtered signals to the greatest extent possible, thereby significantly improving the signal-to-noise ratio. The acquired force, vibration, and acoustic emission signal data are loaded from the host computer's hard drive. The loaded signal data undergoes preprocessing operations to remove DC offset, low-frequency drift, and high-frequency noise to ensure data accuracy and reliability. Specifically, for removing DC offset, the mean of each signal is calculated and subtracted from the signal to remove the DC offset. For filtering low-frequency drift and high-frequency noise, a bandpass filter is used to filter the signal, removing low-frequency drift and high-frequency noise. The length L of the inverse filter is set according to the signal characteristics and requirements. The parameters of the inverse filter are initialized with random values, and the objective function is set to minimum entropy. The output generated during the contact process between the tool and the workpiece is then calculated. For impulse signals, a longer inverse filter can capture more signal features, but it also increases computational complexity. Minimum entropy means minimizing the entropy of the output signal, i.e., the information in the signal is most concentrated. Normalized kurtosis is used as the objective function to optimize the inverse filter. Through iterative calculation, the parameters of the inverse filter are adjusted to maximize the kurtosis, thereby achieving the minimum entropy condition. The optimized inverse filter is then applied to the original signal to determine the optimal inverse filter matrix. Time-domain and frequency-domain analyses are performed on the filtered signal. The waveform and impulse component changes are observed through time-domain analysis, and the spectrum and frequency components changes are observed through frequency-domain analysis. The time-domain waveforms and frequency-domain spectra of the signals before and after filtering are compared to verify whether the impulse component is significantly enhanced and whether the signal-to-noise ratio is improved. The time-domain waveforms and frequency-domain spectra of the original signal and the filtered signal are compared to visually demonstrate the filtering effect, which helps to confirm whether the minimum entropy deconvolution filtering has successfully extracted the impulse component in the signal and weakened its adverse effects.

[0073] Furthermore, the formula for calculating the impact signal generated during the contact process between the tool and the workpiece is as follows:

[0074]

[0075] Where y(n) is the impact signal generated during the contact process between the tool and the workpiece, f(n) is the inverse filter, x(n) is the original acquired signal, L is the length of the inverse filter, * is the convolution operation, f in f(l) refers to the convolution / inverse filter set in the minimum entropy deconvolution filtering method, which is equivalent to a convolution kernel, n is the nth sample in the signal time series, and l is the lth coefficient in the convolution kernel;

[0076] Since a larger kurtosis value (the fourth-order center moment of the normalized data) results in a smaller entropy value, the normalized kurtosis is chosen as the objective function in minimum entropy deconvolution filtering (MED), and its expression is:

[0077]

[0078] When the formula is When the maximum value is reached, the extreme value condition is satisfied. Then it can be determined that f(l) is the optimal inverse filter, and the following equation can be obtained by combining formula (1):

[0079]

[0080] in, For normalized kurtosis, y(i) represents the i-th sample in the time series of the output impulse signal, x, y, and f functions are the original acquired signal, the output impulse signal, and the convolution / inverse filter (convolution kernel) set in the minimum entropy deconvolution filtering (MED) method, respectively. n and l represent the n-th or l-th sample in the signal time series, p represents the p-th coefficient in the convolution kernel, and i, n, l, p represent the i-th, n, l, and p-th values ​​in the x, y, and f functions, respectively.

[0081] Formula (3) can be written in matrix form:

[0082] f = A -1 ×b; (4)

[0083] The optimal inverse filter matrix is ​​finally determined by iterative calculation using formula (4).

[0084] Step 3: Feature extraction is performed on the signal filtered by minimum entropy deconvolution. A total of 95 time-domain and frequency-domain statistical features reflecting turning damage are extracted, including time-domain statistical indicators such as peak value, mean, root mean square value, standard deviation, kurtosis, and skewness, and frequency-domain statistical indicators such as spectral centroid and mean square frequency. The signal data processed by minimum entropy deconvolution filtering is imported into signal analysis software. Turning damage feature values ​​are extracted from the signals filtered by different sensors, which are time-domain and frequency-domain statistical features respectively. These provide input variables for subsequent pattern recognition or machine learning models. Time-domain analysis is performed on the signals filtered by different sensors, and peak value, mean, root mean square value, standard deviation, and skewness are extracted from the time-domain signals. Eleven time-domain statistical indicators are included, such as peak value, kurtosis, skewness, peak-to-peak value, peak factor, impulse factor, waveform factor, and margin factor. Among them, peak value is the maximum value in the signal, reflecting the signal's impact intensity; mean is the average value of the signal, used to assess the overall signal level; root mean square (RMS) is the square root of the average of the squared values ​​of the signal, used to measure the signal's energy level; standard deviation is the square root of the average of the squared differences between the signal values ​​and their mean, reflecting the signal's dispersion; kurtosis is the ratio of the fourth central moment of the signal to the square of the variance, used to describe the signal's sharpness; skewness is the ratio of the third central moment of the signal to the cube root of the variance, used to describe the signal's symmetry; and peak value... Peak value is the difference between the maximum and minimum values ​​in the signal. Peak factor is the ratio of peak value to root mean square (RMS) value, used to evaluate the signal's impulse characteristics. Impulse factor is the ratio of peak value to average value, reflecting the signal's impulse characteristics. Waveform factor is the ratio of RMS value to average value, describing the signal's waveform characteristics. Margin factor is the ratio of peak value to absolute average value, used to evaluate the signal's dynamic range. A Fast Fourier Transform (FFT) is performed on the time-domain signal to obtain its spectral information. From this, eight frequency-domain statistical indicators are extracted, including spectral centroid, mean square frequency, RMS frequency, frequency variance, spectral peak value, power spectral entropy, spectral kurtosis, and spectral skewness. The spectral centroid is the ratio of the peak value to the average value of the signal's frequency spectrum. The weighted average frequency reflects the center position of the signal in the frequency domain. The mean square frequency is the weighted average frequency of the square of the signal spectrum, used to measure the frequency distribution characteristics of the signal. The root mean square frequency is the root mean square value of the spectrum. The frequency variance is used to measure the dispersion of the spectrum distribution. The peak value of the spectrum is used to determine the maximum value in the spectrum and its corresponding frequency. The power spectral entropy is used to measure the complexity of the spectrum. The spectral kurtosis is a statistical measure to describe the sharpness of the peak value of the spectrum. The spectral skewness is used to measure the asymmetry of the spectrum distribution. The time domain and frequency domain features extracted from each sensor are integrated to form a feature vector, and a total of 95 statistical features are extracted. Among them, the force signal includes forces in the X, Y, and Z directions.

[0085] Step 4: Kernel Principal Component Analysis (KPCA) is used for feature dimensionality reduction and fusion. Min-max standardization and a radial basis function (RBF) kernel are applied to map the original data to a high-dimensional feature space. The kernel matrix of the feature matrix is ​​calculated, and eigenvalue decomposition is performed. The top principal components with cumulative contribution rates reaching a certain threshold are selected, and the original data is projected onto the selected principal components to obtain the dimensionality-reduced data. The extracted 95 statistical features are organized into a feature matrix, with each row representing a sample and each column representing a feature. Min-max standardization is applied to the feature matrix to scale the values ​​of each feature to the [0,1] interval, eliminating the influence of different features on their dimensions and ensuring the data are of the same magnitude. Kernel function calculation is then performed, using the radial basis function (RBF). The RBF kernel, also known as the Gaussian kernel, is suitable for mapping nonlinear data. Using the selected RBF kernel, the kernel function value between each sample point in the original dataset is calculated, forming a kernel matrix. Each element of the kernel matrix represents the value between two sample points. The kernel function value is obtained by subtracting the mean of the corresponding row and column from each element in the kernel matrix to center the kernel matrix, ensuring the correct mapping of data in high-dimensional space, eliminating bias in the data, and making subsequent feature decomposition more accurate. Eigenvalue decomposition is performed on the centered kernel matrix to obtain eigenvalues ​​and corresponding eigenvectors. The eigenvalues ​​represent the amount of information contained in each principal component, and the eigenvectors represent the direction of the principal components. Based on the magnitude of the eigenvalues, several (first few) principal components with a cumulative contribution rate reaching a certain threshold (such as 95% or 99%) are selected. These principal components can retain the information of the original data to the greatest extent. The original data is projected onto the selected principal components. By multiplying the kernel matrix with the eigenvectors of the selected principal components, the dimensionality-reduced data is obtained. The dimensionality-reduced data not only retains the main variation information of the original data, but also greatly reduces the dimensionality, thereby effectively reducing redundant information, improving the computational efficiency of subsequent models, and outputting the dimensionality-reduced data. The dimensionality-reduced eigenvectors are the final fused feature set, which will be used as the input of subsequent machine learning models.

[0086] Furthermore, the formula for calculating the radial basis function (RBF) kernel function is as follows:

[0087]

[0088] Among them, K jk z is the inner product of sample j and sample k in the high-order feature space. j and z k It is the vector representation of two samples in the original feature space, where σ is the kernel width parameter;

[0089] Step 5: The least squares support vector machine (LSSVM) method optimized by genetic algorithm is used for pattern recognition to identify the material removal patterns of hard and brittle materials during ultra-precision turning and to determine the state of machining damage.

[0090] Example 2, as Figures 1 to 4 As shown, based on Embodiment 1, the present invention provides a technical solution: Preferably, in step 5, the process for determining the state of processing damage is as follows:

[0091] The reduced feature vector is used as input to match the material removal mode and damage state label corresponding to hard and brittle materials to obtain the reduced feature set, which is used to train and test the LSSVM model. The feature set is divided into training set and test set, with a ratio of 70% training set and 30% test set.

[0092] Furthermore, the removal modes of hard and brittle materials in turning can be divided into two main categories: brittle removal and plastic removal. Plastic removal specifically refers to the phenomenon in ultra-precision turning operations where the material undergoes plastic deformation under the cutting action of the tool, resulting in continuous chips rather than fracture. This process involves the rearrangement of the atomic or molecular structure inside the material and usually occurs below the yield strength threshold of the material. In contrast, brittle removal describes the removal of hard and brittle materials under the action of the tool through the initiation, propagation, and eventual breakage of cracks. When the cutting depth of the tool is large or the applied cutting force exceeds a certain limit, hard and brittle materials tend to undergo brittle fracture, resulting in the generation of a large number of fragments and the formation of cracks. Under this removal mode, the machined surface often exhibits rough characteristics, accompanied by significant cracks and breakage marks, which seriously affects the surface quality. Therefore, by using advanced signal monitoring technology to accurately determine or predict the material removal mode of hard and brittle materials in ultra-precision machining, it is possible to effectively assess whether there is damage to the surface of the machined material.

[0093] Furthermore, the damage status is categorized into: no damage / near-no damage, low-level damage, moderate-level damage, and high-level damage. No damage / near-no damage: The material surface is completely removed by plastic removal, with no surface or subsurface cracks. Low-level damage: The material surface has minimal microscopic damage, with plastic removal as the primary removal mode, accompanied by a small number of microcracks. Moderate-level damage: The surface has obvious cracks and micro-pits, with a mixture of brittle and plastic removal as the material removal mode, and localized brittle fracture. High-level damage: The material surface has obvious deep cracks, spalling, or severe microstructural damage, with brittle fracture as the primary removal mode.

[0094] A least squares support vector machine (LSSVM) classification model is constructed. A radial basis function (RBF) kernel function is selected, and a loss function is defined. The LSSVM classification model learns the classification boundary of the data by solving a least squares optimization problem with regularization terms. By minimizing the squared loss function, the problem is transformed into solving a system of linear equations, significantly improving computational efficiency. A genetic algorithm (GA) is used to optimize the hyperparameters (such as regularization parameters and kernel width) of the LSSVM classification model. The GA generates multiple candidate solutions through selection, crossover, and mutation operations, and evaluates the fitness of each solution (based on the classification accuracy obtained from cross-validation). The global search mechanism of the GA effectively avoids the shortcomings of manual adjustment and reduces the risk of finding local optima. The adaptability of the GA allows it to flexibly adjust parameters on different datasets, thereby improving the model's adaptability and generalization ability. Through iterative optimization using the GA, the optimal combination of hyperparameters is found, thus improving the LSSVM classification model. To assess the classification performance of the model, after determining the optimal hyperparameters, the least squares support vector machine (LSSVM) classification model is trained using the training set data. During training, the model learns how to map different input features to the corresponding output categories (i.e., plasticity removal or brittleness removal). After training, the model is evaluated using the test set. The model performance is measured by calculating indicators such as classification accuracy, recall, precision, and F1-score. The confusion matrix can be used to visualize the model's classification results. The sample data to be identified is input into the trained least squares support vector machine (LSSVM) classification model to obtain the recognition results. The expected value of the feature values ​​under normal turning conditions is set, and the feature values ​​of the statistical features in the dimensionality-reduced feature set are combined to calculate the turning damage recognition coefficient. Then, the damage state of the hard and brittle materials is analyzed. Based on the recognition results and the damage state analysis results, the material removal mode of the hard and brittle materials in the ultra-precision turning process is determined. Based on the identified material removal mode, combined with the processing parameters and process conditions, the damage state in the processing process is determined.

[0095] Furthermore, the expression for calculating the turning damage identification coefficient is as follows:

[0096]

[0097] Where D is the turning damage identification coefficient, representing the damage state, g t Let be the t-th eigenvalue, a statistical feature of the dimensionality-reduced feature set; M be the total number of features; B be the expected value of the eigenvalue under normal turning conditions, determined through experiments or historical data analysis; P be the limiting value, representing the maximum permissible deviation between the eigenvalue and the baseline value, used to adjust the sensitivity of the exponential function; and D ranges from 0 to 1, with values ​​closer to 1 indicating more severe damage. This represents the sum of the square roots of all eigenvalues, used to capture the overall trend of the eigenvalues. It is a reciprocal function, combined with the natural logarithm, used to adjust for the effects of the summation result, making the formula more robust to outliers. It is a sigmoid function used to map the deviation between an eigenvalue and a benchmark value to a range of 0 and 1. The closer the value is to 1, the greater the deviation from the benchmark value. When D approaches B, it approaches 0, indicating no damage or very low damage. When D is significantly greater than B, D is close to 1, indicating a high degree of damage.

[0098] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for monitoring damage during ultra-precision turning of hard and brittle materials based on multi-sensor fusion, characterized in that, Includes the following steps: Step 1: Build a multi-sensor signal acquisition device on an ultra-precision turning machine tool to capture different types of sensor signals generated during the turning process in real time; Step 2 involves performing minimum entropy deconvolution filtering on the acquired sensor signals for signal preprocessing. The signal preprocessing process is as follows: Step 201: Load the collected force signal, vibration signal and acoustic emission signal data from the host computer hard disk, and perform preprocessing operations on the loaded signal data to remove DC offset, filter low frequency drift and high frequency noise. Step 202: Set the length L of the inverse filter according to the characteristics and requirements of the signal, initialize the parameters of the inverse filter with random values, and set the objective function to minimum entropy, and then calculate the impact signal generated during the contact process between the tool and the workpiece. Step 203: Use normalized kurtosis as the objective function to optimize the inverse filter. Through iterative calculation, adjust the parameters of the inverse filter to maximize the kurtosis, thereby achieving the minimum entropy condition. Then apply the optimized inverse filter to the original signal to determine the optimal inverse filter matrix. Step 204: Perform time-domain and frequency-domain analysis on the filtered signal. Observe the changes in the waveform and impulse components of the signal through time-domain analysis, and observe the changes in the spectrum and frequency components of the signal through frequency-domain analysis. Compare the time-domain waveform and frequency-domain spectrum of the signal before and after filtering to verify whether the impulse components are significantly enhanced and whether the signal-to-noise ratio is improved. Step 3: Extract features from the signal after minimum entropy deconvolution filtering, extracting a total of 95 time-domain and frequency-domain statistical features reflecting turning damage; Step 4: Use kernel principal component analysis (KPCA) to perform feature dimensionality reduction and fusion, mapping the original data to a high-dimensional feature space, calculating the kernel matrix of the feature matrix, and performing eigenvalue decomposition to obtain the dimensionality-reduced data. The process of obtaining the dimensionality-reduced data is as follows: Step 401: The 95 extracted statistical features are organized into a feature matrix, with each row representing a sample and each column representing a feature. The feature matrix is ​​then subjected to Min-Max standardization to scale the value of each feature to the range of [0,1], eliminating the influence of different dimensions between features and ensuring that the data are of the same magnitude. Kernel function calculation is then performed. Step 402: Select the radial basis function as the kernel function. Using the selected radial basis function, calculate the kernel function value between each sample point in the original dataset to form a kernel matrix. Each element of the kernel matrix is ​​the kernel function value between two sample points. Step 403: Subtract the mean of the corresponding row and column from each element in the kernel matrix to achieve the centering of the kernel matrix; Step 404: Perform eigenvalue decomposition on the centered kernel matrix to obtain eigenvalues ​​and corresponding eigenvectors. Based on the magnitude of the eigenvalues, select multiple principal components whose cumulative contribution rate reaches the corresponding threshold. Step 405: Project the original data onto the selected principal components. By multiplying the kernel matrix with the eigenvectors of the selected principal components, the dimensionality-reduced data is obtained and the dimensionality-reduced data is output. The dimensionality-reduced eigenvectors are the final fusion feature set. Step 5: A least-squares support vector machine method optimized by a genetic algorithm is used for pattern recognition to identify the material removal patterns of hard and brittle materials during ultra-precision turning and to determine the machining damage state. The process for determining the machining damage state is as follows: Step 501: The reduced feature vector is used as input to match the material removal mode and damage state label corresponding to the hard and brittle material to obtain the reduced feature set. The feature set is divided into training set and test set. The removal mode of hard and brittle material in turning is divided into two categories: brittle removal and plastic removal. The damage state is divided into: no damage / near no damage, low degree of damage, medium degree of damage and high degree of damage. Step 502: Construct a least squares support vector machine classification model, select the radial basis function, and define the loss function; Step 503: Use a genetic algorithm to optimize the hyperparameters of the least squares support vector machine classification model. The genetic algorithm generates multiple candidate solutions through selection, crossover and mutation operations, and evaluates the fitness of each solution. The optimization process is iterated until the optimal combination of hyperparameters is found. Step 504: After determining the optimal hyperparameters, train the least squares support vector machine classification model using the training set data. After training, evaluate the model using the test set. Step 505: Input the sample data to be identified into the trained least squares support vector machine classification model to obtain the identification result; Step 506: Set the expected value of the feature value under normal turning conditions, combine the feature value of the statistical feature in the feature set after dimensionality reduction, calculate the turning damage identification coefficient, and then analyze the damage state of hard and brittle materials. Step 507: Based on the identification results and damage state analysis results, determine the material removal mode of the hard and brittle material in the ultra-precision turning process, and determine the damage state in the processing process based on the identified material removal mode and the processing parameters and process conditions. The formula for calculating the turning damage identification coefficient is as follows: ； in, The turning damage identification coefficient represents the damage state. For the first The eigenvalues ​​are statistical features of the feature set after dimensionality reduction. The total number of features, The expected value of the eigenvalues ​​under normal turning conditions. The limit value represents the maximum permissible deviation between the characteristic value and the reference value. The value ranges from 0 to 1, and the closer the value is to 1, the more severe the damage.

2. The method for monitoring damage during ultra-precision turning of hard and brittle materials based on multi-sensor fusion according to claim 1, characterized in that: In step 1, the process of capturing sensor signals during the turning process is as follows: Step 101: Install acoustic emission sensors, force sensors, and vibration sensors on the cutting edge of the ultra-precision turning machine tool near the material processing area to capture acoustic characteristics, cutting force, and mechanical vibration information during the processing, and install a signal preprocessing device, including a current amplifier and a filter. Step 102: Connect the signal line output by the sensor to the data acquisition card to ensure that the data acquisition card is correctly connected and compatible with the host computer. Configure the sampling rate and resolution of the data acquisition card and set the input channels of the data acquisition card to correspond to the signal input of each sensor. Step 103: Start the data acquisition card to continuously acquire the output signals of each sensor during the turning process. After the sensor signals are processed by the current amplifier and filter, they are transmitted to the data acquisition card of the data acquisition module for analog-to-digital conversion. Step 104: Store the acquired digital signals on the host computer's hard drive and install the corresponding signal processing software on the host computer to receive, store, and preliminarily process the acquired digital signals. At the same time, perform a preliminary check on the data from the perspectives of signal integrity, synchronization, and timestamp verification.

3. The method for monitoring damage during ultra-precision turning of hard and brittle materials based on multi-sensor fusion according to claim 1, characterized in that: The formula for calculating the impact signal generated during the contact process between the cutting tool and the workpiece is as follows: （1）； in, This refers to the impact signal generated during the contact process between the cutting tool and the workpiece. It is an inverse filter. The original acquired signal, The length of the inverse filter. For convolution operations, In This refers to the convolution / inverse filter set in the minimum entropy deconvolution filtering method, which is equivalent to a convolution kernel. The first in the signal time series One sample, The first in the convolution kernel One coefficient; Since a larger kurtosis value corresponds to a smaller entropy value, normalized kurtosis is chosen as the objective function in minimum entropy deconvolution filtering, and its expression is: （2）； When the formula is When the maximum value is reached, the extreme value condition is satisfied. Then it can be determined For the optimal inverse filter, the following equation can be obtained by combining formula (1): （3）； in, To normalize kurtosis, Represents the first in the time series of the output impulse signal One sample, , , The functions respectively represent the original acquired signal, the output impulse signal, and the convolution / inverse filter set in the minimum entropy deconvolution filtering method. and Represents the first in the signal time series or One sample, Represents the first in the convolution kernel One coefficient, , , , Represent , , The first in the function , , , One value; Formula (3) can be written in matrix form: （4）； The optimal inverse filter matrix is ​​finally determined by iterative calculation using formula (4).

4. The method for monitoring damage during ultra-precision turning of hard and brittle materials based on multi-sensor fusion according to claim 3, characterized in that: In step 3, the extraction process of the time-domain and frequency-domain statistical features of turning damage is as follows: Step 301: Import the signal data processed by minimum entropy deconvolution filtering into the signal analysis software, and extract the turning damage feature values ​​of the filtered signals from different sensors, which are time domain and frequency domain statistical features respectively. Step 302: Perform time-domain analysis on the filtered signals from different sensors and extract 11 time-domain statistical indicators from the time-domain signals, including peak value, mean, root mean square value, standard deviation, kurtosis, skewness, peak-to-peak value, peak factor, impulse factor, waveform factor, and margin factor. Step 303: Perform a fast Fourier transform on the time-domain signal to obtain the signal spectrum information, and extract eight frequency domain statistical indicators, including the spectral centroid, mean square frequency, root mean square frequency, frequency variance, spectral peak value, power spectral entropy, spectral kurtosis, and spectral skewness. Step 304: Integrate the time-domain and frequency-domain features extracted from each sensor to form a feature vector, and extract a total of 95 statistical features, including force signals in the X, Y, and Z directions.

5. The method for monitoring damage during ultra-precision turning of hard and brittle materials based on multi-sensor fusion according to claim 4, characterized in that: The formula for calculating the radial basis function kernel function is as follows: ； in, It is a sample and samples The inner product in the high-level feature space, and It is the vector representation of two samples in the original feature space. It is the kernel width parameter.

Images