Method for monitoring wear of CFRP machining tool based on acoustic emission signal

Abstract

The application discloses a CFRP machining tool wear monitoring method based on acoustic emission signals, and relates to the technical field of composite material machining monitoring. The method comprises the following steps: collecting original acoustic emission signals in a CFRP machining process and completing preliminary pretreatment; performing time-frequency domain joint transformation on the pretreated signals to generate acoustic emission joint feature maps containing time domain waveforms, frequency spectrum graphs and time-frequency spectrum graphs; extracting multi-dimensional feature parameters and establishing a nonlinear mapping relationship model with a tool wear state; calculating a quantitative wear index through multi-level feature fusion and weight self-adaptive adjustment in the model; comparing the index with a preset wear grade threshold to determine a wear grade; and outputting a wear state description and maintenance decision suggestion in combination with cutting process parameters. The method can completely retain wear-related signal features, dynamically adapt feature weights, accurately quantitatively evaluate CFRP machining tool wear, and improve the accuracy and monitoring reliability of wear grade determination.

Description

Technical Field

[0001] This invention belongs to the field of composite material processing monitoring technology, specifically a method for monitoring tool wear in CFRP machining based on acoustic emission signals. Background Technology

[0002] In the processing of carbon fiber reinforced composite materials, tool wear monitoring mainly relies on acoustic emission signals. Existing monitoring methods often use single time-domain analysis, frequency-domain analysis, or time-frequency analysis to process acoustic emission signals, which can only obtain a single type of signal feature spectrum, resulting in significant limitations in feature extraction dimensions. Current tool wear condition assessments mostly use fixed-weight feature fusion methods, relying on simple mapping relationships to complete wear condition determination, and cannot adjust the feature matching logic for different wear stages.

[0003] Single-dimensional signal analysis cannot fully retain the detailed information related to tool wear in acoustic emission signals. Feature representation is one-sided. Fixed-weight feature fusion methods cannot adapt to the dynamic changes in tool wear during CFRP machining. The quantitative assessment of wear status is not accurate enough, and there is a large deviation in the determination of wear level.

[0004] It is necessary to realize the joint time-frequency domain transformation of acoustic emission signals to generate a joint feature map containing time-domain waveforms, spectrum diagrams, and time-spectrum diagrams. In the nonlinear mapping relationship model, multi-level feature fusion and weight adaptive adjustment are used to obtain quantitative wear indicators, thereby improving the accuracy and reliability of tool wear monitoring. Summary of the Invention

[0005] The present invention aims to solve at least one of the technical problems existing in the prior art.

[0006] Therefore, this invention proposes a method for monitoring tool wear in CFRP machining based on acoustic emission signals, including: The raw acoustic emission signals generated during the processing of carbon fiber reinforced composite materials are collected to form a raw acoustic emission signal stream. The raw acoustic emission signal stream is then preprocessed to obtain a preprocessed acoustic emission signal. The preprocessed acoustic emission signal is subjected to a joint time-frequency domain transformation to generate a joint acoustic emission feature map containing a time-domain waveform, a spectrum, and a time-spectrum map; Multidimensional feature parameters are extracted from the acoustic emission joint feature map, and a nonlinear mapping relationship model between the multidimensional feature parameters and the tool wear state is established. The multidimensional feature parameters are then input into the nonlinear mapping relationship model. In the nonlinear mapping relationship model, a quantitative wear index reflecting the current comprehensive wear degree of the tool is calculated through multi-level feature fusion and adaptive weight adjustment. The quantitative wear index is compared with the preset wear level threshold range to determine the specific wear level of the current tool. Based on the specific wear level and combined with the current cutting process parameters, the corresponding monitoring conclusions are output, including a description of the wear status and maintenance decision recommendations.

[0007] Further, preliminary preprocessing is performed on the original acoustic emission signal stream to obtain a preprocessed acoustic emission signal, including: The preprocessing includes gain amplification, bandpass filtering, and background noise suppression; The original acoustic emission signal stream is input into a programmable gain amplifier, and the signal amplification factor is automatically adjusted according to the dynamic range of the signal to obtain a pre-amplified acoustic emission signal. The pre-amplified acoustic emission signal is input into an analog bandpass filter with a preset cutoff frequency to filter out signal components below the lower cutoff frequency and above the upper cutoff frequency, thereby obtaining a bandpass filtered acoustic emission signal. Adaptive background noise suppression processing is performed on the bandpass-filtered acoustic emission signal. The processing includes estimating the spectral characteristics of the current ambient noise, generating a noise reference template on the spectral characteristics, and subtracting the noise component corresponding to the noise reference template from the bandpass-filtered acoustic emission signal to obtain the preprocessed acoustic emission signal.

[0008] Further, the preprocessed acoustic emission signal undergoes a joint time-frequency domain transformation to generate a joint acoustic emission feature map containing a time-domain waveform, a spectrogram, and a time-spectrum map, including: The preprocessed acoustic emission signal is digitized at a high sampling rate to obtain a digitized acoustic emission signal sequence; The time-domain waveform is obtained by directly plotting the digitized acoustic emission signal sequence. The fast Fourier transform algorithm is applied to the digitized acoustic emission signal sequence to calculate the frequency domain energy distribution of the signal and plot the spectrum. Apply continuous wavelet transform to the digitized acoustic emission signal sequence, calculate the energy density distribution of the signal in the time-frequency plane, and plot the time-frequency spectrum. The time-domain waveform, the spectrogram, and the time-spectrum are aligned and synthesized according to a preset layout to generate the acoustic emission joint feature map.

[0009] Furthermore, multidimensional feature parameters are extracted from the acoustic emission joint feature map, including: The multidimensional feature parameters include signal energy distribution, frequency band energy ratio, signal amplitude in a specific frequency band, time-frequency ridge slope, and waveform complexity index. On the time-domain waveform of the acoustic emission joint feature map, the root mean square value of the signal within a fixed time window is calculated to obtain the signal energy distribution; On the spectrum of the acoustic emission joint feature map, the ratio of the sum of the energy of the pre-divided sub-bands to the total energy of the entire frequency band is calculated to obtain the frequency band energy proportion; On the spectrum of the acoustic emission joint feature map, locate the specific frequency band associated with the physical phenomena of carbon fiber fracture and matrix cracking, read the maximum value of the signal amplitude within the specific frequency band, and obtain the signal amplitude of the specific frequency band. On the time-frequency spectrum of the acoustic emission joint feature map, identify the time-frequency ridge line with concentrated energy, calculate the slope of the time-frequency ridge line on the time-frequency plane as a function of time, and obtain the slope of the time-frequency ridge line. On the time-domain waveform of the acoustic emission joint feature map, the sum of the number of local maxima and minima of the waveform is calculated, and the ratio of this sum to the total length of the waveform is used as the waveform complexity index.

[0010] Furthermore, a nonlinear mapping model is established between the multidimensional feature parameters and the tool wear state, including: Acquire a set of multidimensional feature parameters corresponding to the acoustic emission signals collected by the tool at different wear stages during historical processing, as well as the corresponding real tool wear labels obtained through offline measurement; Construct a feedforward neural network structure and use the multidimensional feature parameters as the input nodes of the feedforward neural network; The feedforward neural network is trained under supervision using the multidimensional feature parameter sample set and the real tool wear label. The supervised training adjusts the internal connection weights and node biases of the feedforward neural network through the backpropagation algorithm. During training, when the error between the predicted wear amount of the feedforward neural network on the validation set samples and the actual tool wear amount label is less than the preset convergence threshold, training is stopped and the network parameters are fixed at this time to form a trained nonlinear mapping relationship model.

[0011] Furthermore, in the aforementioned nonlinear mapping model, through multi-level feature fusion and adaptive weight adjustment, a quantitative wear index reflecting the current overall wear level of the tool is calculated, including: The multidimensional feature parameters extracted in real time are input into the trained nonlinear mapping model; In the hidden layer of the nonlinear mapping model, the multidimensional feature parameters are transformed by a nonlinear activation function, and each hidden layer node outputs an intermediate feature that integrates some of the input features. The intermediate features output from the last hidden layer of the nonlinear mapping model are passed to the output layer nodes. The output layer node performs a linear weighted sum of all the intermediate features of the input and adds a bias to finally output a continuous value, which is the quantized wear index.

[0012] Furthermore, the quantitative wear index is compared with a preset wear level threshold range to determine the specific wear level of the current tool, including: The wear levels include the initial wear stage, the normal wear stage, and the rapid wear stage; Based on the tool design life and the allowable range of the process, the threshold range of quantitative wear indexes is set in advance to divide the initial wear stage, normal wear stage and rapid wear stage; Read the value of the currently calculated quantitative wear index; The value of the quantitative wear index is compared with the threshold range of the quantitative wear index corresponding to the initial wear stage. If the value of the quantitative wear index falls within the threshold range of the quantitative wear index corresponding to the initial wear stage, the specific wear level is determined to be the initial wear stage. If the value of the quantitative wear index does not fall within the threshold range of the quantitative wear index corresponding to the initial wear stage, it continues to be compared with the threshold range of the quantitative wear index corresponding to the normal wear stage. If it falls within the threshold range, the specific wear level is determined to be the normal wear stage. If the value of the quantitative wear index does not fall within the threshold range of the quantitative wear index corresponding to the normal wear stage, then the specific wear level is determined to be the rapid wear stage.

[0013] Furthermore, based on the specific wear level and the current cutting process parameters, corresponding monitoring conclusions are output, including: The current cutting process parameters, including spindle speed, feed rate, and depth of cut, are read from the real-time interface of the CNC machining system. Establish a standard condition description library and a standard maintenance decision suggestion library containing different wear levels and different combinations of cutting process parameters; Using the specific wear level and the current cutting process parameters as joint query conditions, the closest entry is matched in the standard state description library to obtain the corresponding wear state description text; Using the specific wear level and the current cutting process parameters as joint query conditions, the closest entry is matched in the standard maintenance decision suggestion library to obtain the corresponding maintenance decision suggestion text; The wear condition description text and the maintenance decision recommendation text are combined to generate the monitoring conclusion.

[0014] Furthermore, the establishment of a standard condition description library and a standard maintenance decision suggestion library containing different wear levels and different combinations of cutting process parameters includes: We collected experimental data on the entire process of a cutting tool transitioning from the initial wear stage to the rapid wear stage under various typical combinations of cutting process parameters. The organization's field experts analyzed the experimental data and, for each specific combination of wear level and cutting process parameters, manually wrote corresponding standard condition description texts and standard maintenance decision recommendation texts. A mapping relationship is established between the wear level, the cutting process parameter combination, the standard condition description text, and the standard maintenance decision suggestion text, and stored in the form of a structured data table to form the standard condition description library and the standard maintenance decision suggestion library.

[0015] Furthermore, the method also includes a step of monitoring the online updating of the model: During the machining monitoring process, the multidimensional feature parameters, the quantified wear index, and the actual tool flank wear amount measured offline are periodically recorded to form an online incremental dataset; When the accumulated data volume of the online incremental dataset reaches a preset update threshold, the online incremental dataset is merged with the historical training dataset to form an expanded training dataset. Using the extended training dataset, the parameters of the nonlinear mapping model are retrained and fine-tuned to generate an updated nonlinear mapping model. The current nonlinear mapping relationship model used for real-time monitoring is replaced with the updated nonlinear mapping relationship model to complete an online update of the monitoring model.

[0016] Compared with the prior art, the beneficial effects of the present invention are: Performing a joint time-frequency domain transformation on the preprocessed acoustic emission signal can simultaneously generate a joint acoustic emission feature map consisting of a time-domain waveform, a spectrum, and a time-frequency spectrum. This can completely preserve various wear-related information of the acoustic emission signal in the time domain, frequency domain, and time-frequency coupled state, realizing the synchronous representation of signal features in multiple analysis dimensions. This avoids the limitation of a single analysis dimension being able to extract only local signal features, allowing the extracted multi-dimensional feature parameters to completely cover the signal change details corresponding to tool wear during CFRP machining. It ensures that the correlation characteristics between feature parameters and tool wear state remain highly consistent, improving the comprehensiveness and detail of feature representation, and ensuring that the feature extraction stage can completely capture the signal changes caused by tool wear.

[0017] In the nonlinear mapping model constructed by multidimensional feature parameters and tool wear state, a multi-level feature fusion and weight adaptive adjustment processing method is adopted. According to the contribution of different feature parameters to the representation of tool wear state, the corresponding weights can be dynamically allocated and adjusted, realizing the orderly integration and effective correlation of feature information at different levels. This weakens the interference of redundant features on wear assessment results, strengthens the representation role of core wear features, and directly calculates a quantitative wear index that reflects the current comprehensive wear degree of the tool. This overcomes the problem of insufficient feature adaptability in the fixed weight fusion mode, making the quantitative wear index highly matched with the actual wear state of the tool, improving the accuracy of wear level determination, and making the tool wear monitoring results more consistent with the actual working conditions of CFRP machining. This provides an accurate quantitative reference for the output of wear state description and maintenance decision suggestions. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the steps of the CFRP machining tool wear monitoring method based on acoustic emission signals described in this invention. Figure 2 This is a flowchart of the preliminary preprocessing steps; Figure 3 A flowchart for generating a joint acoustic emission feature map using the joint time-frequency domain transformation. Detailed Implementation

[0019] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and 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.

[0020] See Figure 1 This invention provides a method for monitoring tool wear in CFRP machining based on acoustic emission signals. The specific method includes: Raw acoustic emission signals acquired by sensors during the processing of carbon fiber reinforced composite materials are collected, forming a continuous or segmented raw acoustic emission signal stream. Preliminary preprocessing is performed on this raw acoustic emission signal stream to improve signal quality and obtain a preprocessed acoustic emission signal. A joint time-frequency domain transformation is performed on the preprocessed acoustic emission signal, which simultaneously generates a time-domain waveform, a spectrogram, and a time-spectrum graph. These images are then synthesized into a joint acoustic emission feature map according to a preset layout. Multiple feature parameters containing time-domain, frequency-domain, and time-frequency-domain information are extracted from the joint acoustic emission feature map to form a multidimensional feature parameter set. A nonlinear mapping model between the multidimensional feature parameters and the tool wear state is established, and the real-time extracted multidimensional feature parameters are input into this model. During the internal calculation of this nonlinear mapping model, a quantitative wear index value reflecting the current comprehensive wear degree of the tool is finally calculated through multi-level feature fusion and adaptive weight adjustment. The calculated quantitative wear index value is compared with a preset wear level threshold range, and the specific wear level of the tool is determined based on the comparison result. Based on the determined specific wear level and combined with the current cutting process parameters read in real time from the machine tool CNC system, the monitoring conclusion is output, which includes a description of the wear status and specific maintenance decision suggestions.

[0021] In one embodiment of the present invention, see [reference] Figure 2 The initial preprocessing of the raw acoustic emission signal stream includes three main stages: gain amplification, bandpass filtering, and background noise suppression. A programmable gain amplifier receives the raw acoustic emission signal stream acquired by the sensor. Based on the real-time changes in the peak-to-peak value of the signal in the raw acoustic emission signal stream, the programmable gain amplifier automatically adjusts the signal amplification factor, outputting the pre-amplified acoustic emission signal. In specific implementations, an analog bandpass filter receives the pre-amplified acoustic emission signal. The lower cutoff frequency of the analog bandpass filter is set to 20 kHz to filter out low-frequency interference such as mechanical vibration, and the upper cutoff frequency is set to 1 MHz to filter out high-frequency components such as electronic noise, outputting the bandpass-filtered acoustic emission signal.

[0022] The adaptive background noise suppression module operates on the bandpass-filtered acoustic emission signal. First, it acquires a background noise signal while the tool is idling and not in contact with the workpiece, and estimates the spectral characteristics of the background noise signal using a Fast Fourier Transform (FFT). Based on these spectral characteristics, the module generates a noise reference template that matches the spectral characteristics of the background noise. In practice, generating the noise reference template involves averaging and smoothing the background noise spectrum. During the cutting process, the spectral component corresponding to the noise reference template is subtracted from the real-time acquired bandpass-filtered acoustic emission signal spectrum. Optionally, the subtraction operation is performed point-by-point in the frequency domain. Finally, the processed spectrum is converted back to the time domain using an inverse Fourier transform to obtain the preprocessed acoustic emission signal with suppressed background noise.

[0023] In some embodiments, the adaptive background noise suppression processing module employs spectral subtraction. The core of spectral subtraction is estimating and subtracting the power spectrum of the background noise. The noise reference template corresponds to the estimated background noise power spectrum. ,in Indicates frequency, The power spectrum is obtained by averaging the power spectra of multiple pure noise signal frames. For each frame of the bandpass-filtered acoustic emission signal to be processed, its power spectrum is calculated. The power spectrum of the noise-reduced signal It can be calculated using the formula: in: It is the power spectrum estimation of the denoised signal. It is the power spectrum of the noisy signal. It is the background noise power spectrum. The oversubtraction factor is used to control the amount of noise subtraction. The lower limit factor is used to prevent excessive subtraction from generating musical noise. It can be understood that formula (1) ensures that the power spectrum after noise reduction is not lower than a small proportion of the original signal power spectrum, in order to avoid introducing significant signal distortion. The over-subtraction factor can be understood as... Spectrum lower limit factor The specific values ​​need to be adjusted according to the noise characteristics of the actual processing environment.

[0024] In some embodiments, the gain adjustment strategy of the programmable gain amplifier is based on the dynamic range of the signal. In a specific implementation, the programmable gain amplifier monitors the amplitude of the input signal in real time. When the amplitude of the input signal is continuously lower than a preset threshold, the programmable gain amplifier automatically increases the gain factor. When the amplitude of the input signal exceeds a specific proportion of the analog-to-digital converter's range, the programmable gain amplifier automatically decreases the gain factor. In a specific implementation, the passband range of the analog bandpass filter is set to 50 kHz to 800 kHz. Optionally, the acoustic emission signal after bandpass filtering undergoes an anti-aliasing low-pass filter before entering the adaptive background noise suppression processing module.

[0025] In one embodiment of the present invention, see [reference] Figure 3 The preprocessed acoustic emission signal undergoes a joint time-frequency domain transformation. In practice, high-sampling-rate digitization is achieved using a 16-bit analog-to-digital converter (ADC) with a sampling rate of 5 MHz. The ADC converts the preprocessed acoustic emission signal into a digitized acoustic emission signal sequence, which is a set of time- and amplitude-discrete signal sample points. The time-domain waveform is generated by plotting time on the x-axis and the amplitude of the digitized acoustic emission signal sequence on the y-axis in an image coordinate system, ultimately forming a time-domain waveform that displays the instantaneous change of signal amplitude over time. In practice, the generation of the spectrum requires applying a Fast Fourier Transform (FFT) algorithm to the digitized acoustic emission signal sequence. The FFT transforms the digitized acoustic emission signal sequence from the time domain to the frequency domain, calculates the amplitude or energy of the signal at each frequency component, and plots it with frequency on the x-axis and the calculated amplitude or energy on the y-axis to obtain the spectrum. The spectrum reflects the distribution of the signal's energy along the frequency axis.

[0026] Generating a time-frequency spectrum requires applying continuous wavelet transform to the digitized acoustic emission signal sequence. In practice, the continuous wavelet transform performs scaling and translation operations on the digitized acoustic emission signal sequence using a selected mother wavelet function, calculating the wavelet coefficients at different scales and time points. The squared modulus of the wavelet coefficients represents the energy density of the signal in the time-frequency plane. By plotting time on the horizontal axis and the frequency corresponding to the scale on the vertical axis, and displaying the squared modulus of the wavelet coefficients in grayscale or color-coded form, a time-frequency spectrum can be generated. The time-frequency spectrum can simultaneously display the frequency components of the signal and their changes over time. In some embodiments, the mother wavelet function used for the continuous wavelet transform is the Morlet wavelet, which is a Gaussian function modulated by a complex exponential function. Its mathematical expression is: in: Indicated by time Morlet mother wavelet function as the independent variable It is a normalization constant used to ensure that the energy of the wavelet function is 1. is a natural constant, and the number 5 is the center frequency parameter of the Morlet wavelet. It can be understood that the Morlet wavelet has good time-frequency localization characteristics, making it suitable for analyzing transient non-stationary signals such as acoustic emission signals.

[0027] The time-domain waveform, spectrogram, and time-spectrum graph are aligned and synthesized according to a preset layout. This preset layout can be vertical, with the time-domain waveform at the top, the spectrogram in the middle, and the time-spectrum graph at the bottom. The time axes of the three graphs are strictly aligned horizontally, and the frequency axes are aligned vertically. The synthesized graph forms a joint acoustic emission feature map that comprehensively displays the time-domain, frequency-domain, and time-frequency-domain characteristics of the signal. In some embodiments, the generation of the joint acoustic emission feature map is completed by a dedicated signal processing software module. This module receives the digitized acoustic emission signal sequence, sequentially calls the waveform plotting function, the fast Fourier transform function, and the continuous wavelet transform function, and combines the graphical outputs of the three functions into a single image object according to a preset layout. Optionally, the joint acoustic emission feature map is saved in real-time as an image file or transmitted in memory as an image data stream. Optionally, the vertical axis of the spectrogram uses a logarithmic coordinate system to better represent the differences in energy levels across different frequency bands. It can be understood that the joint acoustic emission feature map provides a unified graphical data source for subsequent extraction of feature parameters from multiple perspectives.

[0028] In one embodiment of the present invention, multidimensional feature parameters are extracted from the acoustic emission joint feature map. These extracted multidimensional feature parameters include signal energy distribution, frequency band energy proportion, signal amplitude in a specific frequency band, time-frequency ridge slope, and waveform complexity index. On the time-domain waveform of the acoustic emission joint feature map, the root mean square value of the signal within a set time window is calculated, and this is used as the signal energy distribution parameter. On the spectrogram, the spectrum is pre-divided into multiple sub-bands, and the ratio of the sum of the energy in each sub-band to the total energy of the entire frequency band is calculated, resulting in a set of frequency band energy proportion parameters. On the spectrogram, specific frequency bands associated with processing physical phenomena such as carbon fiber fracture and matrix cracking are located, and the maximum signal amplitude within these specific frequency bands is read, serving as the signal amplitude parameter for that specific frequency band. On the time-spectrum, time-frequency ridges with concentrated energy are identified, and the slope of these ridges on the time-frequency plane as a function of time is calculated, yielding the time-frequency ridge slope parameter. In the time domain waveform, calculate the total number of local maxima and local minima of the waveform, and use the ratio of this number to the total length of the waveform as the waveform complexity index.

[0029] Establishing a nonlinear mapping model between multidimensional feature parameters and tool wear state requires acquiring historical machining data. Acoustic emission signals collected during historical machining processes at different tool wear stages are processed to obtain a corresponding set of multidimensional feature parameter samples. Offline measurement tools are used to obtain the actual tool wear label corresponding to each sample. A feedforward neural network structure is constructed, with the number of input layer nodes matching the dimension of the multidimensional feature parameters. The acquired multidimensional feature parameter sample set and the actual tool wear labels are used for supervised training of the feedforward neural network. The training process employs the backpropagation algorithm, iteratively adjusting the internal connection weights and node biases of the feedforward neural network. A validation set is set during training. When the error between the network's predicted wear amount and the actual tool wear label for the validation set samples reaches a preset convergence threshold, training is stopped and the network parameters are fixed, forming the trained nonlinear mapping model.

[0030] In this nonlinear mapping model, the quantified wear index is calculated by inputting real-time extracted multidimensional feature parameters into the trained nonlinear mapping model. Nodes in the hidden layers transform the input features using nonlinear activation functions, with each hidden layer node outputting an intermediate feature that incorporates some of the input features. All intermediate features output from the last hidden layer are passed to the output layer nodes. The output layer nodes perform a linear weighted sum of all the input intermediate features and add a bias value, ultimately outputting a continuous numerical value, which is the quantified wear index.

[0031] Multidimensional feature parameters are extracted from the acoustic emission joint feature map. These parameters include signal energy distribution, frequency band energy proportion, signal amplitude in a specific frequency band, time-frequency ridge slope, and waveform complexity index. In practice, the signal energy distribution is calculated on the time-domain waveform of the acoustic emission joint feature map. A time window with a length of 1024 sampling points is selected, and the sum of squares of the amplitudes of all data points within this time window is calculated. The ratio of the sum of squares to the length of the time window is then calculated to obtain the signal energy distribution value within that time window. The formula for calculating the signal energy distribution value is: Energy = Average Amplitude Squared. In practice, the frequency band energy proportion is calculated on the spectrum of the acoustic emission joint feature map. The frequency axis of the spectrum is pre-divided into three sub-bands. The sub-band division is shown in Table 1. Table 1: Sub-band Division Table for Frequency Band Energy Proportion Calculation

[0032] For each sub-band, the spectral energy integral within its frequency range is calculated. The energy integral of each sub-band is compared with the total energy integral of the entire frequency band; the ratio is the frequency band energy proportion of that sub-band, thus obtaining a set of frequency band energy proportion parameters containing three values. The signal amplitude of a specific frequency band is located on the spectrum. In specific implementations, the specific frequency band associated with carbon fiber fracture is located between 150kHz and 250kHz, and the specific frequency band associated with matrix cracking is located between 300kHz and 400kHz. Within these two specific frequency bands, the maximum spectral amplitude is found, and these two maximum values ​​are used as the signal amplitude parameters for that specific frequency band. The slope of the time-frequency ridge is identified on the time-spectrum graph of the acoustic emission joint feature map. The concentrated energy bands on the time-spectrum graph are identified as time-frequency ridges. In practice, a local maximum energy tracking algorithm is used to extract the time-frequency ridges. The slope of the time-frequency ridge is obtained by calculating the ratio of the frequency difference to the time difference between two consecutive points on the ridge, and then averaging the ratios over the entire ridge sequence. The slope of the time-frequency ridge reflects how quickly the dominant frequency component of the signal changes over time. The waveform complexity index is calculated on the time-domain waveform. It iterates through the time-domain waveform data points, identifying all points that satisfy the local maximum condition (the amplitude of this point is greater than the amplitudes of the two points before and after it) and the local minimum condition (the amplitude of this point is less than the amplitudes of the two points before and after it). The total number N of these points is counted. The waveform complexity index is the ratio of the total number N to the total length L of the waveform data points, i.e., waveform complexity index = N / L.

[0033] A nonlinear mapping model between multidimensional feature parameters and tool wear state was established. In the specific implementation, the process of acquiring historical machining data involves conducting CFRP milling experiments under various cutting parameters. The flank wear value (VB) is used as the label for the actual tool wear. Each experiment records the acoustic emission signal and the corresponding flank wear value (VB) at different times from when the tool is new to when it fails. After processing, multiple sets of paired data of multidimensional feature parameter samples and actual tool wear labels are obtained. The constructed feedforward neural network structure includes one input layer, two hidden layers, and one output layer. The input layer has 7 nodes, corresponding to seven feature parameters: signal energy distribution, frequency band energy ratio (3), signal amplitude in a specific frequency band (2), and time-frequency ridge slope. The first hidden layer contains 10 nodes, the second hidden layer contains 5 nodes, and the output layer contains 1 node for outputting the quantified wear index. The feedforward neural network was trained under supervision using a collected set of multidimensional feature parameters and real tool wear labels. During training, the sample data was randomly divided into training, validation, and test sets. The backpropagation algorithm was used to minimize the mean squared error between the network's predicted output and the real tool wear labels. The Adam algorithm was used as the optimizer. Training was stopped when the mean squared error loss on the validation set stopped decreasing after 20 consecutive training epochs, and the network connection weights and node bias parameters were fixed, forming a trained nonlinear mapping model.

[0034] The quantified wear index is calculated in the trained nonlinear mapping model. In practice, the seven feature parameter values ​​extracted in real time are input into the nonlinear mapping model. In the first hidden layer of the nonlinear mapping model, each node receives a weighted input from all nodes in the input layer. The weighted input is processed by the nonlinear activation function ReLU. The expression of the ReLU function is: in: This represents the weighted sum of the node's inputs. This represents the output of the ReLU activation function. After the ReLU transformation, each node in the first hidden layer outputs an intermediate feature value that incorporates some input features. The intermediate feature values ​​from the first hidden layer are passed to the second hidden layer. In the second hidden layer, nodes undergo similar calculations and transformations, outputting second-layer intermediate feature values. These second-layer intermediate feature values ​​are then passed to the output layer nodes. The output layer nodes perform a linear weighted sum of all the second-layer intermediate feature values. The formula for the linear weighted sum is: in: This represents the weighted sum of the output layer nodes. This represents the weight connecting the i-th node in the second hidden layer to the node in the output layer. This represents the output value of the i-th node in the second hidden layer. This represents the bias value of the output layer nodes. No nonlinear activation function is applied to the linear weighted sum calculation result; the final output value... This refers to the quantified wear index. It can be understood that the entire calculation process achieves a nonlinear mapping from multidimensional feature parameters to a single wear value. The nonlinear activation function of the hidden layer nodes is crucial for realizing this complex mapping relationship. Optionally, during model application, the calculated quantified wear index can be further normalized to ensure its value range is between 0 and 1. Optionally, before receiving the feature parameters, the input layer of the nonlinear mapping model will perform standardization processing on the feature parameters, consistent with the training phase.

[0035] In one embodiment of the present invention, a quantified wear index is compared with a preset wear level threshold range to determine the specific wear level. In specific implementations, the preset wear levels include an initial wear stage, a normal wear stage, and a rapid wear stage. The preset wear level threshold range is set based on the tool manufacturer's lifespan data and the allowable flank wear limit of the process. The quantified wear index threshold range is defined as follows: the threshold range for the initial wear stage is greater than or equal to 0 and less than 0.2; the threshold range for the normal wear stage is greater than or equal to 0.2 and less than 0.8; and the threshold range for the rapid wear stage is greater than or equal to 0.8. The value of the quantified wear index calculated and output by the nonlinear mapping model is read. In specific implementations, the value of the quantified wear index is a real number between 0 and 1. The value of the quantified wear index is compared with the threshold range for the initial wear stage. If the value of the quantified wear index is greater than or equal to 0 and less than 0.2, the current wear level of the tool is determined to be the initial wear stage. If the value of the quantitative wear index does not fall within the threshold range corresponding to the initial wear stage, the value of the quantitative wear index is further compared with the threshold range corresponding to the normal wear stage. If the value of the quantitative wear index is greater than or equal to 0.2 and less than 0.8, the specific wear level is determined to be the normal wear stage. If the value of the quantitative wear index is greater than or equal to 0.8, the specific wear level is determined to be the rapid wear stage.

[0036] Based on the specific wear level and combined with the current cutting process parameters, the monitoring conclusion is output. The current cutting process parameters are read from the real-time data interface of the CNC machining system. In specific implementation, the spindle speed, feed rate, and depth of cut are read from the machine tool controller via OPCUA or MTConnect protocol. The system has established a standard status description library and a standard maintenance decision suggestion library. The standard status description library stores the standard status description text corresponding to different wear levels and different combinations of cutting process parameters, and the standard maintenance decision suggestion library stores the corresponding standard maintenance decision suggestion text. Using the real-time determined specific wear level and the read current cutting process parameters as joint query conditions, a match is made in the standard status description library. The standard status description library adopts a segmented interval matching strategy. For example, the query conditions "specific wear level = normal wear stage, spindle speed range = [8000, 12000] rpm, feed rate range = [800, 1200] mm / min, depth of cut range = [0.8, 1.2] mm" match the closest entry and obtain the corresponding wear status description text. Using the same joint query conditions, a match is made in the standard maintenance decision suggestion library to obtain the corresponding maintenance decision suggestion text. In specific implementation, the wear status description text and the maintenance decision suggestion text are merged and a complete monitoring conclusion is generated by string concatenation.

[0037] The establishment of the standard condition description library and the standard maintenance decision recommendation library was accomplished through the following steps: Experimental data was collected on the entire process of a tool transitioning from the initial wear stage to the rapid wear stage under various typical combinations of cutting process parameters. The experimental data included acoustic emission characteristics recorded at different wear levels, process parameters, and offline measurements of the flank wear. Domain experts analyzed the experimental data, and for each specific wear level and cutting process parameter combination, they manually wrote corresponding standard condition description texts and standard maintenance decision recommendation texts. An example of a standard condition description text is: "The tool is in the normal wear stage. Under the current parameters of 12000 rpm spindle speed, 1000 mm / min feed rate, and 1.0 mm depth of cut, the wear progression rate is stable." An example of a standard maintenance decision recommendation text is: "It is recommended to continue using the tool and re-inspect it after machining 20 parts." A mapping relationship was established between the wear level, cutting process parameter combination, standard condition description text, and standard maintenance decision recommendation text, and stored in the form of a structured data table to form the standard condition description library and the standard maintenance decision recommendation library. In some embodiments, the standard condition description library and the standard maintenance decision recommendation library are stored in relational database tables, see Table 2: Table 2: Standard Maintenance Decision Recommendation Database

[0038] It is understood that Table 2 is merely an example, and the actual database contains many more entries with combinations of process parameters. During query matching, if no identical range of the current process parameter is found in the database, the nearest neighbor matching algorithm is used. This algorithm calculates the Euclidean distance between the current parameter and the center point of the parameter range of each entry in the database, and selects the entry with the smallest distance as the matching result. The formula for calculating the Euclidean distance is: in: Represents Euclidean distance. Indicates the current spindle speed. This indicates the center value of the spindle speed range for each entry in the library table. Indicates the current feed rate. This represents the center value of the feed rate range for each entry in the library table. Indicates the current cutting depth. This represents the center value of the cutting depth range for each entry in the database table. It's understood that the nearest neighbor matching algorithm ensures that a reasonable suggestion is output even when the process parameters do not precisely match the preset range in the database table. Optionally, the standard condition description library and the standard maintenance decision suggestion library support adding, deleting, and modifying entries through database management tools.

[0039] In one embodiment of the present invention, during continuous machining monitoring, the multidimensional feature parameters used in each monitoring session, the quantitative wear index calculated by the model, and the actual tool flank wear obtained through offline measurement after the tool has finished being used are periodically recorded. These three data points are correlated to form a data record, which is continuously accumulated to form an online incremental dataset. When the amount of data in the online incremental dataset reaches a preset update threshold, the online incremental dataset is merged with the historical training dataset previously used to train the model to form an extended training dataset. Using this extended training dataset, the parameters of the currently deployed nonlinear mapping relationship model are retrained and fine-tuned to generate an updated nonlinear mapping relationship model. The newly trained and updated nonlinear mapping relationship model replaces the current model used for real-time monitoring, completing one online update of the monitoring model.

[0040] The online update step of the monitoring model is executed during the machining monitoring process. In specific implementation, periodic recording is performed based on each complete tool machining stroke or a fixed time interval. Within each monitoring cycle, the system records the multidimensional feature parameter vector extracted from the acoustic emission joint feature spectrum and the quantified wear index value calculated by the nonlinear mapping relationship model. After the tool is used and removed, the actual wear width VB value of the tool flank is measured offline using a tool microscope. The multidimensional feature parameter vector, quantified wear index value, actual tool flank wear VB value, and corresponding process parameters and timestamps are associated to form a complete data record, which is stored in the incremental data buffer designated by the system. The set of all recorded data forms the online incremental dataset. When the amount of data in the online incremental dataset reaches the preset update threshold, which is defined by the number of data records, the model update process is triggered when the online incremental dataset accumulates 200 new data records.

[0041] After triggering the model update process, the online incremental dataset is merged with the historical training dataset. The historical training dataset is the complete dataset used in the previous training of the nonlinear mapping model. The merging operation appends 200 records from the online incremental dataset to the end of the historical training dataset, forming an expanded training dataset with a larger sample size and potentially broader coverage. The parameters of the currently deployed nonlinear mapping model are retrained and fine-tuned using the expanded training dataset. The retraining and fine-tuning process uses the connection weights and node bias parameters of the currently deployed nonlinear mapping model as initial values, employing the same loss function and optimizer as the initial model training. A new round of training iterations is performed on the expanded training dataset. The number of iterations in this new round is set to a fixed value, such as 50 iterations, or it stops early when the model's performance on the validation set reserved from the expanded training dataset no longer improves. An updated nonlinear mapping model is generated. The updated nonlinear mapping model refers to the nonlinear mapping model with new connection weights and node bias parameters after retraining and fine-tuning.

[0042] The online update of the monitoring model replaces the current nonlinear mapping relationship model used for real-time monitoring with an updated nonlinear mapping relationship model. In specific implementations, the model replacement operation is performed during the idle period of the system monitoring service or during planned downtime. The new model file is loaded into the monitoring system through file replacement or hot loading of memory parameters, replacing the original model file. The prediction module in the monitoring service is then restarted, enabling subsequent real-time wear monitoring to be calculated based on the updated nonlinear mapping relationship model. In some embodiments, the online model update process includes a verification step. This verification step uses a portion of the data in the online incremental dataset or an independent test set to evaluate the prediction accuracy of the updated nonlinear mapping relationship model. The final model replacement operation is only performed if the evaluation metric is better than or equal to the current model. The evaluation metric can be the mean absolute error or the coefficient of determination R². Optionally, the online incremental dataset is cleared after being used for model updates to prepare for receiving new data. Simultaneously, the merged extended training dataset will be used as a new historical training dataset for the next model update. It can be understood that the online model update step enables the nonlinear mapping relationship model to adapt to the effects of tool batch differences, material batch differences, or slow drift in process conditions. Optionally, the preset update threshold can be dynamically adjusted based on the model performance degradation rate in actual applications. The adjustment is based on the error trend between the recent predicted value of the quantified wear index and the actual tool flank wear VB value measured offline. When the average error continues to exceed a set threshold, the preset update threshold can be automatically reduced, for example, from 200 to 100, to trigger model updates more frequently. The formula for calculating the error trend can be: in: This represents the average absolute error within the time window t. This indicates the number of data points used for calculation within the time window t. This represents the actual tool flank wear measured offline at the i-th data point. This represents the predicted wear amount after dimensional transformation of the corresponding model's predicted quantitative wear index value. This represents absolute value operations. It can be understood that by monitoring error trends and dynamically adjusting the update threshold, the model update strategy can be made more adaptive. In some embodiments, the online model update step can be set to a manually triggered mode, whereby the operator manually initiates the update process through the user interface when deemed necessary.

[0043] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for monitoring tool wear in CFRP machining based on acoustic emission signals, characterized in that, include: The raw acoustic emission signals generated during the processing of carbon fiber reinforced composite materials are collected to form a raw acoustic emission signal stream. The raw acoustic emission signal stream is then preprocessed to obtain a preprocessed acoustic emission signal. The preprocessed acoustic emission signal is subjected to a joint time-frequency domain transformation to generate a joint acoustic emission feature map containing a time-domain waveform, a spectrum, and a time-spectrum map; Multidimensional feature parameters are extracted from the acoustic emission joint feature map, and a nonlinear mapping relationship model between the multidimensional feature parameters and the tool wear state is established. The multidimensional feature parameters are then input into the nonlinear mapping relationship model. In the nonlinear mapping relationship model, a quantitative wear index reflecting the current comprehensive wear degree of the tool is calculated through multi-level feature fusion and adaptive weight adjustment. The quantitative wear index is compared with the preset wear level threshold range to determine the specific wear level of the current tool. Based on the specific wear level and combined with the current cutting process parameters, the corresponding monitoring conclusions are output, including a description of the wear status and maintenance decision recommendations.

2. The method for monitoring CFRP machining tool wear based on acoustic emission signals according to claim 1, characterized in that, The original acoustic emission signal stream is subjected to preliminary preprocessing to obtain a preprocessed acoustic emission signal, including: The preprocessing includes gain amplification, bandpass filtering, and background noise suppression; The original acoustic emission signal stream is input into a programmable gain amplifier, and the signal amplification factor is automatically adjusted according to the dynamic range of the signal to obtain a pre-amplified acoustic emission signal. The pre-amplified acoustic emission signal is input into an analog bandpass filter with a preset cutoff frequency to filter out signal components below the lower cutoff frequency and above the upper cutoff frequency, thereby obtaining a bandpass filtered acoustic emission signal. Adaptive background noise suppression processing is performed on the bandpass-filtered acoustic emission signal. The processing includes estimating the spectral characteristics of the current ambient noise, generating a noise reference template on the spectral characteristics, and subtracting the noise component corresponding to the noise reference template from the bandpass-filtered acoustic emission signal to obtain the preprocessed acoustic emission signal.

3. The method for monitoring CFRP machining tool wear based on acoustic emission signals according to claim 1, characterized in that, The preprocessed acoustic emission signal is subjected to a joint time-frequency domain transformation to generate a joint acoustic emission feature map containing a time-domain waveform, a spectrogram, and a time-spectrum map, including: The preprocessed acoustic emission signal is digitized at a high sampling rate to obtain a digitized acoustic emission signal sequence; The time-domain waveform is obtained by directly plotting the digitized acoustic emission signal sequence. The fast Fourier transform algorithm is applied to the digitized acoustic emission signal sequence to calculate the frequency domain energy distribution of the signal and plot the spectrum. Apply continuous wavelet transform to the digitized acoustic emission signal sequence, calculate the energy density distribution of the signal in the time-frequency plane, and plot the time-frequency spectrum. The time-domain waveform, the spectrogram, and the time-spectrum are aligned and synthesized according to a preset layout to generate the acoustic emission joint feature map.

4. The method for monitoring CFRP machining tool wear based on acoustic emission signals according to claim 1, characterized in that, Extracting multidimensional feature parameters from the acoustic emission joint feature map includes: The multidimensional feature parameters include signal energy distribution, frequency band energy ratio, signal amplitude in a specific frequency band, time-frequency ridge slope, and waveform complexity index. On the time-domain waveform of the acoustic emission joint feature map, the root mean square value of the signal within a fixed time window is calculated to obtain the signal energy distribution; On the spectrum of the acoustic emission joint feature map, the ratio of the sum of the energy of the pre-divided sub-bands to the total energy of the entire frequency band is calculated to obtain the frequency band energy proportion; On the spectrum of the acoustic emission joint feature map, locate the specific frequency band associated with the physical phenomena of carbon fiber fracture and matrix cracking, read the maximum value of the signal amplitude within the specific frequency band, and obtain the signal amplitude of the specific frequency band. On the time-frequency spectrum of the acoustic emission joint feature map, identify the time-frequency ridge line with concentrated energy, calculate the slope of the time-frequency ridge line on the time-frequency plane as a function of time, and obtain the slope of the time-frequency ridge line. On the time-domain waveform of the acoustic emission joint feature map, the sum of the number of local maxima and minima of the waveform is calculated, and the ratio of this sum to the total length of the waveform is used as the waveform complexity index.

5. The method for monitoring CFRP machining tool wear based on acoustic emission signals according to claim 1, characterized in that, Establishing a nonlinear mapping model between the multidimensional feature parameters and the tool wear state includes: Acquire a set of multidimensional feature parameters corresponding to the acoustic emission signals collected by the tool at different wear stages during historical processing, as well as the corresponding real tool wear labels obtained through offline measurement; Construct a feedforward neural network structure and use the multidimensional feature parameters as the input nodes of the feedforward neural network; The feedforward neural network is trained under supervision using the multidimensional feature parameter sample set and the real tool wear label. The supervised training adjusts the internal connection weights and node biases of the feedforward neural network through the backpropagation algorithm. During training, when the error between the predicted wear amount of the feedforward neural network on the validation set samples and the actual tool wear amount label is less than the preset convergence threshold, training is stopped and the network parameters are fixed at this time to form a trained nonlinear mapping relationship model.

6. The method for monitoring CFRP machining tool wear based on acoustic emission signals according to claim 5, characterized in that, In the aforementioned nonlinear mapping model, a quantitative wear index reflecting the current overall wear level of the tool is calculated through multi-level feature fusion and adaptive weight adjustment, including: The multidimensional feature parameters extracted in real time are input into the trained nonlinear mapping model; In the hidden layer of the nonlinear mapping model, the multidimensional feature parameters are transformed by a nonlinear activation function, and each hidden layer node outputs an intermediate feature that integrates some of the input features. The intermediate features output from the last hidden layer of the nonlinear mapping model are passed to the output layer nodes. The output layer node performs a linear weighted sum of all the intermediate features of the input and adds a bias to finally output a continuous value, which is the quantized wear index.

7. The method for monitoring CFRP machining tool wear based on acoustic emission signals according to claim 1, characterized in that, The quantitative wear index is compared with a preset wear level threshold range to determine the specific wear level of the current tool, including: The wear levels include the initial wear stage, the normal wear stage, and the rapid wear stage; Based on the tool design life and the allowable range of the process, the threshold range of quantitative wear indexes is set in advance to divide the initial wear stage, normal wear stage and rapid wear stage; Read the value of the currently calculated quantitative wear index; The value of the quantitative wear index is compared with the threshold range of the quantitative wear index corresponding to the initial wear stage. If the value of the quantitative wear index falls within the threshold range of the quantitative wear index corresponding to the initial wear stage, the specific wear level is determined to be the initial wear stage. If the value of the quantitative wear index does not fall within the threshold range of the quantitative wear index corresponding to the initial wear stage, it continues to be compared with the threshold range of the quantitative wear index corresponding to the normal wear stage. If it falls within the threshold range, the specific wear level is determined to be the normal wear stage. If the value of the quantitative wear index does not fall within the threshold range of the quantitative wear index corresponding to the normal wear stage, then the specific wear level is determined to be the rapid wear stage.

8. The method for monitoring CFRP machining tool wear based on acoustic emission signals according to claim 1, characterized in that, Based on the specific wear level and the current cutting process parameters, the corresponding monitoring conclusions are output, including: The current cutting process parameters, including spindle speed, feed rate, and depth of cut, are read from the real-time interface of the CNC machining system. Establish a standard condition description library and a standard maintenance decision suggestion library containing different wear levels and different combinations of cutting process parameters; Using the specific wear level and the current cutting process parameters as joint query conditions, the closest entry is matched in the standard state description library to obtain the corresponding wear state description text; Using the specific wear level and the current cutting process parameters as joint query conditions, the closest entry is matched in the standard maintenance decision suggestion library to obtain the corresponding maintenance decision suggestion text; The wear condition description text and the maintenance decision recommendation text are combined to generate the monitoring conclusion.

9. The method for monitoring CFRP machining tool wear based on acoustic emission signals according to claim 8, characterized in that, The establishment of a standard condition description library and a standard maintenance decision suggestion library containing different wear levels and different combinations of cutting process parameters includes: We collected experimental data on the entire process of a cutting tool transitioning from the initial wear stage to the rapid wear stage under various typical combinations of cutting process parameters. The organization's field experts analyzed the experimental data and, for each specific combination of wear level and cutting process parameters, manually wrote corresponding standard condition description texts and standard maintenance decision recommendation texts. A mapping relationship is established between the wear level, the cutting process parameter combination, the standard condition description text, and the standard maintenance decision suggestion text, and stored in the form of a structured data table to form the standard condition description library and the standard maintenance decision suggestion library.

10. The method for monitoring CFRP machining tool wear based on acoustic emission signals according to claim 1, characterized in that, The method also includes a step of monitoring online model updates: During the machining monitoring process, the multidimensional feature parameters, the quantified wear index, and the actual tool flank wear amount measured offline are periodically recorded to form an online incremental dataset; When the accumulated data volume of the online incremental dataset reaches a preset update threshold, the online incremental dataset is merged with the historical training dataset to form an expanded training dataset. Using the extended training dataset, the parameters of the nonlinear mapping model are retrained and fine-tuned to generate an updated nonlinear mapping model. The current nonlinear mapping relationship model used for real-time monitoring is replaced with the updated nonlinear mapping relationship model to complete an online update of the monitoring model.

Images