A Rotating Tool Wear Monitoring Method Based on the Feature Fusion Map of Acoustic-Vibration Signals

Abstract

The present invention discloses a method for monitoring the wear of a rotating tool based on a feature fusion map of acoustic vibration signals. The detection signals are subjected to feature extraction in the time domain, frequency domain, waveform domain, and wavelet domain, and the extracted features are correlated with the tool wear value. Features with large correlation coefficients are selected and fused with the intrinsic features of the rotating tool to establish a feature fusion map data set for tool wear; moreover, a model fusion technology is used to establish an integrated neural network architecture with a three-layer network fusion, and the feature fusion map data set is used for training and verification. The integrated neural network established by the present invention has a prediction accuracy of 97.85% for the wear of the rotating tool, and among the prediction results of individual wear categories in the six wear levels of pre-wear, early-middle wear, middle wear, late-middle wear, late wear, and failure, the prediction accuracy of all wear categories except late-middle wear does not reach 0.9, and the prediction accuracy of the remaining wear categories is above 0.95.

Description

Technical Field

[0001] The present invention belongs to the technical field of tool wear state monitoring, and particularly relates to a method for monitoring the wear state of a rotating tool based on a feature fusion graph of acoustic and vibration signals. Background Art

[0002] In the field of machining, CNC machine tools are the core equipment suitable for machining workpieces of different scales and batches, and rotating tools are the core components in the CNC machining process. The mutual cooperation between the two can achieve various types of precise and complex machining processes. As the direct participant in machining that contacts the workpiece, the tool removes the excess material on the workpiece along the specified feed path to form the machined surface. During the entire process of forming the surface to be machined, the flank face of each cutting edge of the rotating tool contacts and rubs against the formed surface of the workpiece, resulting in progressive wear and dulling of the tool. In machining, tool wear is a direct factor affecting the machining surface quality of parts and the service life of machine tools. According to statistics, the downtime caused by tool failure in cutting machining accounts for 20% of the total downtime, and the cost increased due to the tool itself and tool change accounts for 3% - 12% of the total cost. Timely obtaining the tool wear state can effectively avoid workpiece quality problems caused by tool abnormalities, and improve the utilization rate of tools by grinding the worn tools to reduce costs.

[0003] Tool wear state identification is divided into direct method and indirect method from the measurement method. The direct method uses a high-speed camera in combination with machine vision technology to achieve state recognition based on the information on the tool wear image. On the one hand, it is very difficult to exclude the influence of cutting fluid oil stains, workpiece chip splashing, and machining environment light when identifying the tool wear state by the direct method, resulting in a large error in tool wear state recognition. On the other hand, the direct method requires providing a suitable installation space for the image acquisition instrument inside the machine tool, so it is very difficult to apply it to the real machining environment with complex processes. Therefore, it is very difficult to achieve online monitoring of the tool wear state by the direct method. The indirect method is to collect various physical signals generated during the cutting process and analyze the signals related to tool wear to monitor the tool wear state. The indirect method can not only exclude the influence of the harsh cutting environment on data in the direct method, but also has less modification to the mechanical part of the machine tool and is easier to achieve online detection. The physical quantities in the cutting process include signals such as cutting force, vibration, and sound. Selecting appropriate signals for analysis can not only improve the correct rate of wear state recognition, but also reduce the total cost of equipment maintenance, indirectly bringing more profits to the enterprise.

[0004] The force signal is sensitive to the change of contact conditions between the worn tool and the material during the metal cutting process. The shear stress and extrusion stress corresponding to the worn tool change greatly. And research shows that the cutting force is the most sensitive to the worn tool and is the direct cause of vibration signals, sound signals and sound generation signals. However, the price of force sensors is expensive, and it is difficult to popularize them to small and medium-sized processing enterprises or companies. The force sensors of the Swiss company Kistler are unanimously recognized by many research scholars in the processing field for their high-precision measurement of dynamic forces during the cutting process. However, its price is approximately 60% - 80% of that of a million-level CNC machining center. In addition, the fixation of force sensors not only needs to consider the internal space layout of the machine tool itself, but also requires a large-scale transformation of the fixture. Once the size of the workpiece changes greatly, the fixture needs to be redesigned according to the size of the workpiece and the force sensor. To sum up, using force sensors for tool condition monitoring is not the first choice. Compared with the high cost of force sensors, the acceleration sensors used to collect vibration signals are characterized by low price, easy installation and no need for major mechanical transformation of the machine tool body. In addition, research shows that the dynamic component of the cutting force that changes periodically during the cutting process is the main cause of vibration. Therefore, it is reasonable to use acceleration sensors instead of force sensors to monitor the tool wear state. The cutting noise during the tool wear process contains rich cutting information, which can indicate the tool wear state. Experienced processors can judge tool wear, chipping of the tool edge, ablation of the tool teeth and other various abnormal states by ear according to the noise during the cutting process. Research shows that the sound pressure level of the cutting noise gradually increases as the tool wears to breakage, and the influence of the flank wear on the front of the cutting edge on the sound signal is concentrated in the frequency spectrum from 2.75 kHz to 350 kHz.

[0005] At present, for the problem of tool wear state monitoring, the existing research ideas and solutions have the following defects: (1) In terms of hardware, most use force sensors or acceleration sensors. It is difficult to exclude the noise influence of other machine tools in the cutting environment and the vibration influence caused by installing and fixing large-sized workpieces when using a single type of single sensor for monitoring. In addition, once an abnormality occurs during processing and damages one of the sensors, the entire tool state monitoring system will face the problem of paralysis. (2) The indirect method of tool wear state monitoring is mainly feature-driven. Most of them only use the feature screening algorithm of machine learning or the neural network of deep learning to automatically screen features. The former mostly screens features in the time domain or frequency domain of single-channel signals, with relatively large limitations. The latter has problems such as long training time and poor interpretability. (3) Most of the training sets for tool wear state monitoring based on computer vision are mainly the time series diagrams of cutting signals, the frequency spectrum diagrams of FFT, and the time-frequency diagrams of STFT. These signal diagrams only consider the information of the cutting signals themselves and do not consider the intrinsic information of the tool state, making it difficult to intuitively display the highly correlated features related to tool wear. (4) The decision model of tool wear state only relies on a single machine learning algorithm or a single deep learning neural network model. It will be very difficult to achieve high-precision discrimination of tool wear state and low generalization error with less training time and small batch data. Summary of the Invention

[0006] In view of the problems of low accuracy in monitoring the tool wear state using single-channel signals, poor generality and large generalization error in monitoring the tool wear state of rotary tools with a single model, especially slow feature extraction and poor visualization of tool wear signals in the existing online monitoring technology for the wear state of rotary tools, the present invention provides a method for monitoring the wear of rotary tools based on a feature fusion map of acoustic and vibration signals. It is a method that extracts multi-domain tool wear signal features from the time domain, frequency domain, and wavelet domain by performing spectral transformation and wavelet transformation on the time-domain tool wear signal and considering the intrinsic features of the tool; it is a method that efficiently extracts and fuses the acoustic and vibration signal features highly correlated with the flank wear amount of the tool from multi-domain tool wear signal features through the correlation analysis of tool wear signal features and the principal component extraction technology combined with high-dimensional feature compression; it is a method that comprehensively considers the multi-channel signal changes and multi-domain feature characteristics caused by tool wear and realizes efficient visualization through information graph representation; it is a method that uses ensemble learning technology to fuse multiple small and lightweight convolutional neural networks to achieve high-accuracy and low-generalization-error monitoring of the flank wear state of rotary tools; it is a high-performance monitoring method for the wear state of rotary tools that integrates the extraction of key features of tool wear signals, the efficient visualization of signal features highly correlated with the tool wear amount, and the rapid identification of the flank wear state of the tool.

[0007] A method for monitoring the wear of rotary tools based on a feature fusion map of acoustic and vibration signals according to the present invention is as follows:

[0008] Step 1: Set cutting parameters and tool paths.

[0009] Step 2: Build a tool wear test platform for acoustic and vibration signals. The tool wear test platform for acoustic and vibration signals includes an electron microscope, a signal collector, a three-axis acceleration sensor, a sound level meter, and an image acquisition device. The image acquisition device includes a vertical industrial camera and a horizontal industrial camera. The three-axis acceleration sensor is fixed to the spindle of the machine tool, and the sound level meter, the vertical industrial camera, and the horizontal industrial camera are all fixed to the workbench of the machine tool.

[0010] Step 3: Conduct cutting experiments and obtain three-axis acceleration signals, acoustic signals, VB values, and wear images during the cutting process.

[0011] Step 4: Extract time-domain features and waveform features from the acceleration signals in each direction of the three-axis acceleration signal, a total of 11 types of feature values, namely mean, standard deviation, skewness, kurtosis, peak-to-peak value, root mean square, crest factor, waveform factor, impulse factor, margin factor, and energy. Then, extract frequency-domain features from the acceleration signals in each direction of the three-axis acceleration signal, a total of 11 types of feature values, namely average frequency-domain amplitude, center frequency, mean square frequency, frequency variance, frequency amplitude variance, average frequency, median frequency, total power, average power, frequency corresponding to the maximum power, and ratio of low-frequency power to high-frequency power. Next, perform wavelet packet decomposition and wavelet packet reconstruction on the acceleration signals in each direction of the three-axis acceleration signal using wavelet transform to obtain 8 wavelet energy ratios, 8 wavelet energy entropies, 1 wavelet energy spectrum entropy, and at the same time perform singular value decomposition on the three-axis acceleration signal to obtain 1 wavelet packet singular spectrum entropy. After completing the extraction of time-domain, waveform, frequency-domain, and wavelet-domain features, obtain a vibration signal feature matrix from the three-axis acceleration signal. Finally, for the acoustic signal, only extract features in the time domain to obtain an acoustic signal feature matrix.

[0012] Step 5: First, calculate the correlation coefficients between the time-domain features, frequency-domain features, and wavelet-domain features of the three-axis acceleration signal and the average value of the VB values of the flank faces of the three cutting edges. Then, calculate the correlation coefficient between the time-domain features of the acoustic signal and the average value of the VB values of the flank faces of the three cutting edges.

[0013] Step 6: Screen out the features with a correlation coefficient greater than 0.5 with the average value of the VB values of the flank faces of the three cutting edges to obtain a screened feature matrix. Perform a mean normalization operation on each element in the screened feature matrix, and then realize the fusion of the three-axis acceleration signal, acoustic signal features, and the intrinsic features of the rotating tool.

[0014] Step 7: Divide the wear levels of the rotating tool into 6 categories, as shown in the following table. Among them, the wear value in the tool wear value interval refers to the average value of the wear bandwidth of the flank faces of the three cutting edges;

[0015] Wear level Wear state Wear value range Sample quantity 1 Early wear 0-80 143 2 Mid-early wear 80-95 179 3 Mid wear 95-100 105 4 Mid-late wear 100-120 195 5 Late wear 120-150 160 6 Failure >150 163

[0016] Step 8: Establish a feature fusion graph dataset for tool wear.

[0017] Step 9: Use model fusion technology to establish an integrated neural network architecture and conduct training and verification. Specifically: Build an integrated neural network based on the feature fusion graph, mainly composed of Network 1, Network 2, Network 3, and a voting component. Network 1 serves as the upper layer architecture of the integrated neural network, with a structure of 2 convolutional activation layers, 3 fully connected layers, and 1 Softmax classifier. Network 2 serves as the middle layer architecture of the integrated neural network, with a structure of a combined layer of 2 convolutional activation layers and 1 convolutional pooling layer from left to right, a combined layer of 2 convolutional activation layers and 1 convolutional pooling layer, a combined layer of 3 convolutional activation layers and 1 convolutional pooling layer, a combined layer of 3 convolutional activation layers and 1 convolutional pooling layer, a combined layer of 3 convolutional activation layers and 1 convolutional pooling layer, 2 fully connected activation layers with loss, 1 fully connected layer, and 1 Softmax classifier. Network 3 serves as the lower layer architecture of the integrated neural network, with a structure of a combined layer of 1 convolutional activation layer and 1 convolutional pooling layer, a combined layer of 1 convolutional activation layer and 1 convolutional pooling layer, a combined layer of 3 convolutional activation layers and 1 convolutional pooling layer, 2 fully connected activation layers with loss, 1 fully connected activation layer, 1 fully connected layer, and 1 Softmax classifier. Set the loss function and iterator, train the integrated neural network based on the feature fusion graph using the training set, establish the integrated neural network model metrics based on the feature fusion graph, and verify the integrated neural network based on the feature fusion graph on the test set.

[0018] Preferably, Step 3 is specifically as follows: The machine tool drives the rotating tool to process the workpiece according to the set cutting parameters and feed path. During the processing, once the signal curves of the three-axis acceleration sensor and the sound level meter show abnormalities, or after one feed is completed, the spindle of the machine tool stops rotating and rises for reset, the three-axis acceleration sensor and the sound level meter stop signal acquisition, the working platform moves to align the image acquisition device with the rotating tool, the horizontal industrial camera takes images of the flank wear area of the side cutting edge of the rotating tool, measures the VB values corresponding to the three cutting edges, and the vertical industrial camera takes images of the wear of the bottom cutting edge of the rotating tool. Before the next feed, store the three-axis acceleration signal, sound signal, VB value, and wear images. Conduct experiments on 3 rotating tools, and each rotating tool records the three-axis acceleration signal, sound signal, VB value, and wear images corresponding to n = 315 feeds.

[0019] Preferably, Step 6 is specifically as follows:

[0020] Step 1: Feature screening is performed on the time domain, frequency domain, and wavelet domain features of each direction of the three-axis acceleration signal. Six features with a correlation coefficient greater than 0.5 with the average value of the VB values of the three flank faces of the cutting edge are selected and recorded in an n×6 screened feature matrix, where n is the number of tool passes and 6 is the number of features after screening. Then, the elements in the screened feature matrix are normalized to zero mean. Similarly, feature screening is performed on the time domain features of the acoustic signal. Six features with a correlation coefficient greater than 0.5 with the average value of the VB values of the three flank faces of the cutting edge are selected and recorded in an n×6 screened feature matrix, and the elements in this screened feature matrix are also normalized to zero mean.

[0021] Step 2: Calculate the eigenvalue covariance matrix corresponding to each element in the screened feature matrix obtained from each direction of the acceleration signal after normalizing to zero mean. Then, calculate the eigenvalues and eigenvectors of each covariance matrix. Sort the distinct eigenvalues of each covariance matrix from largest to smallest. Set the number of principal components to 3. Analyze the principal component variance contribution rate after PCA processing of the eigenvectors corresponding to each eigenvalue solution. Take the elements in the three eigenvectors with the largest cumulative variance contribution rate as the principal features of the corresponding direction of the acceleration signal.

[0022] Step 3: Sort the eigenvalues corresponding to each element in the screened feature matrix obtained from the acoustic signal after normalizing to zero mean from largest to smallest, and take the first two features as the principal features of the acoustic signal.

[0023] Step 4: Unitize the cutting time according to the number of tool passes. The time corresponding to one tool pass is defined as one cutting unit. Finally, normalize the number of tool passes to obtain the intrinsic features of the rotary tool.

[0024] Step 5: Combine the principal features of each direction of the acceleration signal, the principal features of the acoustic signal, and the intrinsic features of the rotary tool to obtain the fusion features.

[0025] Preferably, Step 8 is specifically as follows: Use circular polar coordinates as the coordinate system of the fusion features. According to the 12 features of the fusion features, 12 polar axes are equally spaced around the origin. The values on each polar axis are restricted between 0 and 1. Take the number of tool passes as the first axis, and the 2nd to 10th axes are sequentially assigned to the 9 principal features of the three-axis acceleration signal. The remaining 2 axes are assigned to the principal features of the acoustic signal. For each tool pass, mark the values of the fusion features on their respective feature polar axes in sequence, and connect these points in sequence to form a closed loop. After establishing the feature fusion diagram for all tool passes, establish a training set and a test set with a ratio of 8:2.

[0026] Preferably, in Step 9, the steps for establishing the integrated neural network model metrics based on the feature fusion diagram are as follows:

[0027] The first-level indicators for constructing an integrated neural network based on a feature fusion graph are TP, FN, FP, and TN. Among them, TP represents the number of positive samples where the true value is positive and the integrated neural network also considers it positive; FN represents the number of positive samples where the true value is positive but the integrated neural network considers it negative; FP represents the number of negative samples where the true value is negative but the integrated neural network considers it positive; TN represents the number of negative samples where the true value is negative and the integrated neural network also considers it negative.

[0028] More preferably, the second-level indicators for constructing an integrated neural network based on a feature fusion graph are Accuracy, Precision, Recall, and Specificity.

[0029] The beneficial effects of the present invention are as follows:

[0030] The present invention constructs a monitoring method for rotary tool wear based on a feature fusion graph of acoustic and vibration signal features. It extracts time-domain, frequency-domain, waveform, and wavelet-domain features from the detection signal, calculates the correlation between the extracted features and the tool wear value, selects the features with large correlation coefficients and fuses them with the intrinsic features of the rotary tool to establish a feature fusion graph dataset for tool wear; moreover, it uses model fusion technology to establish an integrated neural network architecture with three-layer network fusion, and uses the feature fusion graph dataset for training and verification. The prediction accuracy of the integrated neural network established by the present invention for rotary tool wear reaches 97.85%. Among the prediction results of individual wear categories in the 6 wear levels of early wear, mid-early wear, mid-term wear, mid-late wear, late wear, and failure, except that the accuracy of mid-late wear does not reach 0.9, the prediction accuracy of the remaining wear categories is above 0.95. Description of the Drawings

[0031] Figure 1 It is a flowchart of the present invention;

[0032] Figure 2 It is a structural diagram of the acoustic and vibration signal tool wear test platform in the present invention;

[0033] Figure 3 It is a wear curve diagram of the three cutting edges of the first tool C1 in the present invention;

[0034] Figure 4(a), Figure 4(b), and Figure 4(c) are respectively the correlation diagrams of the Y-direction acceleration signal of the first tool C1 in the present invention in the time domain, frequency domain, and wavelet domain features and the average value of the VB value of the flank face of the three cutting edges of the rotary tool;

[0035] Figure 5 It is the feature fusion graph of the present invention;

[0036] Figure 6This is the architecture diagram of the integrated neural network model based on the feature fusion graph in the present invention. Detailed implementation manners

[0037] The present invention will be further described below with reference to the accompanying drawings.

[0038] As Figure 1 shown, a method for monitoring the wear of a rotary tool based on a feature fusion graph of acoustic and vibration signals is as follows:

[0039] Step 1: Set the cutting parameters and the tool path.

[0040] In this embodiment, the workpiece is machined using the cutting parameters in Table 1. The tool material is HSS-Al, the diameter of the rotary tool 1 is 16 mm, the cutting edge length is 48 mm, the tip radius is 1.6 mm, and the number of cutting edges is 3. The workpiece 2 is made of 45# steel and is a cube with dimensions of 100 mm × 100 mm × 80 mm. The tool path is a straight-line cutting along the Y-axis direction in the 100 mm × 100 mm plane. The tool is lifted every 100 mm and fed 0.1 mm in the X-axis direction, and then the cutting is performed in the reverse direction along the Y-axis for the next straight line. This process is repeated, and one tool path contains 5 straight-line cutting paths.

[0041] Table 1 Cutting parameters

[0042]

[0043] Step 2: Build a test platform for tool wear monitoring of acoustic and vibration signals. As Figure 2 shown, the test platform for tool wear monitoring of acoustic and vibration signals includes an electron microscope, a signal collector, a three-axis acceleration sensor 3, a sound level meter 4, and an image acquisition device.

[0044] Machine tool VM-23S(A) is selected. The tool wear is measured using an electron microscope and is equipped with an annular light source. For the convenience of later multi-channel signal analysis, a 16-channel signal collector and a three-axis acceleration sensor 3 are used, and the AWA5636 sound level meter is used for sound signal acquisition.

[0045] Considering the size of the workpiece, two solutions are provided for the fixation of the three-axis acceleration sensor:

[0046] (1) The three-axis acceleration sensor is fixedly connected to a magnetic base through a threaded connection. The magnetic base is fixed on the spindle of the machine tool and is 50 mm away from the bottom surface of the spindle. For the sake of measurement accuracy, try to keep the center distance between the tool tip and the three-axis acceleration sensor unchanged during each cutting process. This fixation method of the three-axis acceleration sensor is applicable to workpieces with a maximum size less than 100 mm.

[0047] (2) The three-axis acceleration sensor is fixedly connected to the magnetic base through a threaded connection. The magnetic base is fixed on the spindle of the machine tool and is 10 mm away from the bottom surface of the spindle.

[0048] The fixing method of the sound level meter is as follows: The front part of the iron universal rod clamps the middle and rear ends of the sound level meter. The universal rod is fixedly connected to another magnetic base through a threaded connection. The magnetic base is fixedly connected to the working platform of the machine tool by magnetic force, and the universal rod can be adjusted in any direction by itself.

[0049] The image acquisition device includes a vertical industrial camera 5 and a horizontal industrial camera 6. As Figure 2 , the lens heads of the two industrial cameras are ring lights, which are used to provide sufficient ambient light for the tool wear part. The vertical industrial camera and the horizontal industrial camera are both fixedly connected to the optical base plate through threaded connections, and the optical base plate is fixedly connected to the working table of the machine tool through trapezoidal nuts. A fixed scale can be used to provide a suitable pixel equivalent value for the industrial camera.

[0050] Connect the 3 transmission lines of the three-axis acceleration sensor to channels 1, 2, and 3 of the signal collector, corresponding to the vibrations in the X, Y, and Z directions respectively. Connect 1 transmission line of the sound level meter to channel 4 of the signal acquisition. The USB transmission lines of the vertical industrial camera 5 and the horizontal industrial camera 6 are directly connected to the laptop computer. Calibrate the sound level meter with a (sound corrector). When the sound pressure value is displayed as 5.89 - 6.66, it indicates that the instrument can be used normally. Select 20480 as the sampling frequency of the signal collector.

[0051] Step 3: Conduct a cutting experiment and obtain the three-axis acceleration signal, sound signal, VB value, and wear image during the cutting process.

[0052] The machine tool is warmed up for 10 - 15 minutes, and then drives the rotary tool to machine the workpiece according to the set cutting parameters and feed path. During the machining process, once the signal curves of the three-axis acceleration sensor and the sound level meter show abnormalities, or after completing one feed (completing the entire feed path), the spindle of the machine tool stops rotating and rises for reset. The three-axis acceleration sensor and the sound level meter stop signal acquisition. The workbench moves to align the image acquisition device with the rotary tool. The horizontal industrial camera takes images of the flank wear area of the side cutting edge of the rotary tool, measures the VB values corresponding to the three cutting edges, and the vertical industrial camera takes images of the wear of the bottom cutting edge of the rotary tool. Before the next feed, the three-axis acceleration signal, sound signal, VB value, and wear image are stored. This process is repeated in a cycle to conduct tests on 3 rotary tools. Each rotary tool records the three-axis acceleration signal, sound signal, VB value, and wear image corresponding to n = 315 feeds. The 3 rotary tools are numbered according to the cutting order, and 3 file directories are established, named C1, C2, and C3 respectively. Taking C1 as an example, the naming rule for the wear image is C1_1_001_005_flute1, where 1 represents the first feed, 001_005 represents the 1st straight line to the 5th straight line cut, and flute1 represents that this wear image corresponds to the 1st cutting edge of the rotary tool. Taking C1 as an example, the naming rule for the VB value file is C1_csv. C1 is the rotary tool number, and C1_csv records the VB value measurement results of the three cutting edges during 315 cuts. The first three values in each row correspond to the VB value measurement results of the three cutting edges after one feed, and the fourth and fifth values in each row record the average and maximum values of the VB values of the three cutting edges. Taking C1 as an example, the naming rule for the three-axis acceleration signal and sound signal file is C1_1_001_005_ch4, where 1 represents the first feed, 001_005 represents the 1st straight line to the 5th straight line cut, and ch4 represents that 4 channels of signals are collected during this cutting process, namely 3 channels of three-axis acceleration signals and 1 channel of sound signal. The first column in this file is the number of points collected, and the remaining columns correspond to the 3 acceleration signal channels and 1 sound signal channel respectively according to the cutting order.

[0053] Step 4: Extract time domain features and waveform features from each direction acceleration signal of the three-way acceleration signal, with a total of 11 types of feature values, namely, mean, standard deviation, skewness, kurtosis, peak-to-peak value, root mean square, crest factor, waveform factor, pulse factor, margin factor, and energy. The calculation formulas for time domain features and waveform features are shown in Table 2. Then, extract frequency domain features from each direction acceleration signal of the three-way acceleration signal, with a total of 11 types of feature values, namely, frequency domain amplitude average, center of gravity frequency, mean square frequency, frequency variance, frequency amplitude variance, average frequency, median frequency, total power, average power, maximum power corresponding frequency, and low-frequency power to high-frequency power ratio. The calculation formula for frequency features is shown in Table 3. Next, the db6 wavelet transform (the number of wavelet packet decomposition layers is 3) is used to perform wavelet packet decomposition and wavelet packet reconstruction on each direction acceleration signal of the three-dimensional acceleration signal, and 8 wavelet energy ratios, 8 wavelet energy entropies, and 1 wavelet energy spectrum entropy are obtained. At the same time, the three-dimensional acceleration signal is singularly decomposed to obtain 1 wavelet packet singular spectrum entropy, so a total of 18 features are extracted in the wavelet domain. After completing the feature extraction in the time domain, waveform, frequency domain, and wavelet domain, the vibration signal feature matrix obtained from the three-dimensional acceleration signal is 3×40 in size. Finally, for the acoustic signal, since it is easily affected by the sound generated by the operation of the machine tool and the sound generated by the air pump, the feature extraction in the frequency domain and wavelet domain is not performed on the acoustic signal, and the feature extraction is only performed in the time domain. The size of the acoustic signal feature matrix is ​​1×11. At this point, the feature extraction of the acoustic vibration signal (including the three-dimensional acceleration signal and the acoustic signal) is completed, and a total of 131 features are extracted from the three-dimensional acceleration signal and the acoustic signal.

[0054] Table 2 Time domain characteristics and waveform characteristics

[0055]

[0056] Among them, x i Represents the three-dimensional acceleration value collected at time i during a tool pass, N S Represents the total number of three-axis acceleration values ​​collected in one pass.

[0057] Table 3 Frequency domain characteristics

[0058]

[0059]

[0060] Where s(k) is the amplitude corresponding to the kth spectral line in the sequence formed by the FFT transformation of the three-dimensional acceleration signal time domain sequence, S 1 ~S 5 The features obtained from FFT transformation, and S 6 ~S 11 is the feature extracted from the power spectrum; N represents the total number of spectral lines after FFT transformation, fk represents the frequency value of the k-th spectral line, P(f) is the power spectral function, w is the frequency value corresponding to the maximum power, and S max is the amplitude corresponding to the frequency w, and s l is the maximum frequency value in the low-frequency range, and s h is the maximum frequency value in the high-frequency range. The low-frequency range is within 5 kHz, and the rest is the high-frequency range.

[0061] Step 5: Calculate the correlation between the three-axis acceleration signal features, the acoustic signal features, and the average value of the VB values of the three cutting-edge flank faces.

[0062] After extracting the data from C1_csv using the Python Pandas module, the time-domain features, frequency-domain features, and wavelet-domain features of the three channels of the three-axis acceleration signal of the rotary tool are combined with the time-domain features of the acoustic signal. A total of 131 features are used to calculate the correlation coefficients with the average value of the VB values of the three cutting-edge flank faces. Specifically, first calculate the correlation coefficients between the time-domain features, frequency-domain features, and wavelet-domain features of the three-axis acceleration signal and the average value of the VB values of the three cutting-edge flank faces, and then calculate the correlation coefficients between the time-domain features of the acoustic signal and the average value of the VB values of the three cutting-edge flank faces. The correlation coefficient calculation formula is as follows:

[0063]

[0064]

[0065]

[0066] A total of 131 features extracted from the three-axis acceleration signal and the acoustic signal are normalized through Equation (1.1). Among them, X i represents the feature vector of the i-th feature that has not been normalized. Its matrix size is 315 rows and 1 column, denoted as (315,1), and F i represents X i after normalization of the feature vector, and x max is the maximum value in X i , and x min is the minimum value in X i , and X min is a vector composed of 131 x min and has the same matrix size as X i ; the covariance of the feature vectors X and Y is calculated through Equation (1.2). The feature vectors X and Y are taken from two different F i , where x j and y j represent the j-th elements in the feature vectors X and Y respectively, and μ Xrepresents the mean value of all elements in the feature vector X, μ Y represents the mean value of all elements in the feature vector Y, and n represents the number of tool feed passes of the rotary tool; Equation (1.3) is used to calculate the correlation coefficient between the feature vectors X and Y, where σ X represents the variance of all elements in the feature vector X, σ Y represents the variance of all elements in the feature vector Y. Only the correlation coefficient diagrams of the Y-direction acceleration signal in the time domain, frequency domain, and wavelet domain and the average value of the VB values of the flank faces of the three cutting edges of the C1 rotary tool are shown in Figures 4(a), 4(b), and 4(c). In Figure 4(a), 0-10 represents 11 time-domain features of the Y-direction acceleration signal, and 11 represents the average value of the VB values of the flank faces of the three cutting edges. In Figure 4(b), 0-10 represents 11 frequency-domain features of the Y-direction acceleration signal, and 11 represents the average value of the VB values of the flank faces of the three cutting edges. In Figure 4(c), 0-17 represents 18 wavelet-domain features of the Y-direction acceleration signal, and 18 represents the average value of the VB values of the flank faces of the three cutting edges. It should be noted that for the present invention, what is actually useful in Figures 4(a), 4(b), and 4(c) is only the correlation coefficient related to the average value of the VB values of the flank faces of the three cutting edges.

[0067] Step 6: Screen out the features with a correlation coefficient greater than 0.5 with the average value of the VB values of the flank faces of the three cutting edges to obtain a screened feature matrix. Perform a mean-zeroing operation on each element in the screened feature matrix, and then use the Python Scikit-Learn module to realize the fusion of the three-direction acceleration signal, acoustic signal features highly correlated with the wear amount, and the intrinsic features of the rotary tool. The specific steps are as follows:

[0068] Step 1: Perform feature screening on the time-domain, frequency-domain, and wavelet-domain features of each direction acceleration signal in the three-direction acceleration signal, and select 6 features with a correlation coefficient greater than 0.5 with the average value of the VB values of the flank faces of the three cutting edges, and record them in an n×6 screened feature matrix, where n is the number of tool feed passes and 6 is the number of screened features. Then, perform a mean-zeroing operation on each element in the screened feature matrix. The formula is as follows:

[0069]

[0070] where, f ij represents the element in the i-th row and j-th column of the screened feature matrix, is the average value of the j-th column elements, and f′ ij is the value after mean-zeroing of each element in the screened feature matrix.

[0071] Similarly, feature screening is performed on the time-domain features of the acoustic signal, and 6 features with a correlation coefficient greater than 0.5 with the average value of the VB values of the three flank faces of the cutting edge are selected and recorded in the n×6 screened feature matrix. Each element in the screened feature matrix is also subjected to the operation of normalizing the mean to 0.

[0072] Step2. Use the Python Scikit-Learn module to calculate the eigenvalue covariance matrix corresponding to each element in the screened feature matrix obtained from the acceleration signal in each direction after normalizing the mean to 0.

[0073]

[0074] Among them, f j represents the matrix recording the jth feature after n tool passes, f k represents the matrix recording the kth feature after n tool passes, f′ ij and f′ ik respectively represent the jth feature and the kth feature of the ith tool pass, C represents the covariance matrix, and cov(f j , f k ) represents the covariance of f j and f k .

[0075] Then calculate the eigenvalues and eigenvectors of each covariance matrix.

[0076] Cp = λp (1.5)

[0077] Among them, λ is the eigenvalue of C, and p is the eigenvector of the finally obtained covariance matrix.

[0078] Sort the different solutions of the eigenvalues of each covariance matrix from large to small, set the number of principal components to 3, perform analysis on the contribution rate of the principal component variance for the eigenvectors corresponding to each eigenvalue solution after PCA processing, and use the elements in the three eigenvectors with the largest cumulative variance contribution rate as the main features of the acceleration signal in the corresponding direction.

[0079] Step3. Sort the eigenvalues corresponding to each element in the screened feature matrix obtained from the acoustic signal after normalizing the mean to 0 from large to small, and take the first two features as the main features of the acoustic signal.

[0080] Step4. Unitize the cutting time according to the number of tool passes. The time corresponding to one tool pass is defined as one cutting unit. Finally, normalize the number of tool passes to obtain the intrinsic features of the rotary tool.

[0081] Step 5: Combine the main features of the acceleration signals in each direction, the main features of the acoustic signals, and the eigenfeatures of the rotating tool to obtain the fusion features. The number of features of the fusion features is 12, 9 features are from the three-axis acceleration signals, 2 features are from the acoustic signals, and 1 feature is the eigenfeature of the rotating tool (the feature representing the number of tool passes).

[0082] Step 7: Classify the tool wear state by comprehensively considering the concavity and convexity, increase and decrease of the curve, and the order of curve fitting, and combine the set wear index. Finally, the wear levels of the rotating tool are divided into 6 categories, as shown in Table 4. Among them, the wear value in the tool wear value interval refers to the average value of the wear bandwidth of the flank faces of the three cutting edges. Taking C1 as an example, the detailed wear curves of the three cutting edges are as Figure 3 。

[0083] Table 4 Wear levels

[0084] Wear level Wear state Wear value range Sample quantity 1 Early wear 0-80 143 2 Mid-early wear 80-95 179 3 Mid wear 95-100 105 4 Mid-late wear 100-120 195 5 Late wear 120-150 160 6 Failure >150 163

[0085] Step 8: Establish the feature fusion map dataset of tool wear

[0086] Take each tool pass as a sample, and each sample corresponds to 12 features. To adapt to the size of the receptive field of the neural network and improve the training speed, a circular polar coordinate with symmetry is selected as the coordinate system of the fusion features. The circular polar coordinate is to establish 12 polar axes equidistantly around the origin according to the 12 features. The values on each polar axis are limited between 0 and 1. Take the number of tool passes as the first axis, and the 2nd to 10th axes are sequentially assigned to the 9 main features of the three-axis acceleration signals, and the remaining 2 axes are assigned to the main features of the acoustic signals. For each tool pass, mark the values of the fusion features on their respective feature polar axes in sequence, and connect these points in sequence to form a closed loop, thus completing the establishment of the feature fusion map of a sample. Establish the feature fusion maps of all samples. Taking C1 as an example, the final feature fusion map is as Figure 5As shown. Then, a training set and a test set are established according to Table 5. Specifically: create file directories Train and Test under the data directory. Each file directory contains 6 subdirectories, and the corresponding serial numbers of the wear levels and a specified number of feature fusion graphs are placed in each subdirectory. The training set is placed in Train, and the test set is placed in Test. The ratio of the training set to the test set is 8:2. Finally, use torchvision.datasets.ImageFolde, torchvision.transform, and torch.utils.data.DataLoader in Pytorch to encapsulate the feature fusion graph data in the file directories Train and Test under the data directory into a dataset that can be trained and tested by a neural network. Different from using traditional STFT and spectrograms for training, using information graphs (feature fusion graphs) for training requires less time, greatly reducing computing resources, and fundamentally solving the problem of slow model training caused by too many model parameters.

[0087] Table 5 Feature Fusion Graph Dataset

[0088] Wear level category C1 C2 C3 Quantity 1 001-034 001-085 001-024 143 2 035-126 086-143 025-053 179 3 127-150 144-214 054-064 105 4 151-211 215-242 064-169 195 5 212-286 243-269 170-227 160 6 287-315 270-315 228-315 163

[0089] Step Nine: Use model fusion technology to establish an integrated neural network architecture and perform training and verification.

[0090] First, refer to the lightweight network framework in CNN to build an integrated neural network based on feature fusion graphs. The entire network framework is as Figure 6 , and the entire network is mainly composed of Network One, Network Two, Network Three, and a voting component; Network One serves as the upper layer architecture of the integrated neural network, with a structure of 2 convolutional activation layers, 3 fully connected layers, and 1 Softmax classifier. Network Two serves as the middle layer architecture of the integrated neural network, with a relatively complex structure. The corresponding structure from left to right is a combination layer of 2 convolutional activation layers and 1 convolutional pooling layer, a combination layer of 2 convolutional activation layers and 1 convolutional pooling layer, a combination layer of 3 convolutional activation layers and 1 convolutional pooling layer, a combination layer of 3 convolutional activation layers and 1 convolutional pooling layer, a combination layer of 3 convolutional activation layers and 1 convolutional pooling layer, 2 fully connected activation layers with losses, 1 fully connected layer, and 1 Sotfmax classifier. Network Three serves as the lower layer architecture of the integrated neural network, with a structure of a combination layer of 1 convolutional activation layer and 1 convolutional pooling layer, a combination layer of 1 convolutional activation layer and 1 convolutional pooling layer, a combination layer of 3 convolutional activation layers and 1 convolutional pooling layer, 2 fully connected activation layers with losses, 1 fully connected activation layer, 1 fully connected layer, and 1 Softmax classifier.

[0091] Among them, the convolutional activation layer is used to implement the convolutional operation to learn the features of the feature fusion graph. The convolutional pooling layer is used to implement the reduction of the feature fusion graph to achieve dimensionality reduction of the features. The fully connected layer in the fully connected activation layer serves as a linear layer close to the output result, and appropriate weights can be obtained through training. The activation layer uses the Relu function. The Relu activation layer is used to avoid affine transformation and enhance the fitting ability by introducing non-linearity. The fully connected activation layer with loss is different from the fully connected activation layer. By introducing loss technology, the values of some neurons are directly set to 0 to suppress overfitting of the model. The voting component is used to implement the fusion of the models of Network One, Network Two, and Network Three, and the soft voting technology is adopted to take the wear level category of the rotary tool with the highest probability as its prediction.

[0092] The expression of the Relu function is as follows

[0093] f(x) = max(x, 0) (1.6) Among them, x represents the value in the output tensor of the previous layer, and f(x) represents the output value after passing through the Relu function, which is also the output value of the current layer.

[0094] The expression of the Softamx function in the Softmax classifier is as follows

[0095]

[0096] Among them, f y is the value in the output tensor of the fully connected layer, and f c is the output result of the fully connected layer for the wear level category c of the rotary tool. C is the total number of categories to be classified, that is, the number of wear level categories of the rotary tool. The output value p(y|x) of the softmax function represents the probability value that the predicted value is the wear level category y of the rotary tool.

[0097] Secondly, set the loss function and iterator. The Adam iterator is selected. The formula of the loss function L is as follows

[0098]

[0099] Among them, v i is the output result of the wear level category i of the rotary tool in the layer before the Softmax classifier, and v j is the output result of the wear level category j of the rotary tool in the layer before the Softmax classifier.

[0100] Then, train the integrated neural network based on the feature fusion graph through the training set, establish the index of the integrated neural network model based on the feature fusion graph, and verify the integrated neural network based on the feature fusion graph on the test set.

[0101] The learning rate of Network 1 in the training parameters is 0.0001, the learning rate of Network 2 is 0.00001, and the learning rate of Network 3 is 0.0001. Early stopping is used during training to alleviate overfitting of the integrated neural network, and the training period is 300.

[0102] The steps to establish the integrated neural network model metrics based on the feature fusion graph are as follows:

[0103] Step 1: Construct the first-level metrics of the integrated neural network based on the feature fusion graph, namely TP, FN, FP, and TN. Among them, TP represents the number of positive samples where the true value is positive and the integrated neural network also considers it positive; FN represents the number of positive samples where the true value is positive but the integrated neural network considers it negative; FP represents the number of negative samples where the true value is negative but the integrated neural network considers it positive; TN represents the number of negative samples where the true value is negative and the integrated neural network also considers it negative.

[0104] Step 2: When dealing with a large amount of data, it is difficult to measure the quality of the integrated neural network model using the first-level metrics. Therefore, construct the second-level metrics of the integrated neural network based on the feature fusion graph, namely Accuracy, Precision, Recall, and Specificity, and the calculations are as follows:

[0105]

[0106] The present invention has obtained good results in predicting 6 wear grade categories on the test set. The prediction accuracy rate for the entire test set has reached 97.85%. The performance of the second-level metrics for predicting a single wear grade category is shown in Table 6. Among them, except that the accuracy rate of the 4th category has not reached 0.9, the prediction accuracy rates of the other categories are all above 0.95.

[0107] Table 6 Performance of the second-level metrics for the integrated neural network's prediction of a single wear grade category on the test set

[0108]

Claims

1. A method for monitoring the wear of a rotating tool based on a feature fusion graph of acoustic and vibration signals, characterized in that: Specifically as follows: Step 1: Set cutting parameters and feed paths; Step 2: Build a test platform for tool wear of acoustic and vibration signals. The test platform for tool wear of acoustic and vibration signals includes an electron microscope, a signal collector, a three-axis acceleration sensor, a sound level meter, and an image acquisition device. The image acquisition device includes a vertical industrial camera and a horizontal industrial camera. The three-axis acceleration sensor is fixed to the spindle of the machine tool, and the sound level meter, the vertical industrial camera, and the horizontal industrial camera are all fixed to the workbench of the machine tool; Step 3: Conduct cutting experiments and obtain the three-axis acceleration signals, sound signals, VB values, and wear images during the cutting process; Step 4: Extract time-domain features and waveform features from each direction of the three-axis acceleration signals, a total of 11 types of feature values, namely mean, standard deviation, skewness, kurtosis, peak-to-peak value, root mean square, crest factor, waveform factor, impulse factor, margin factor, and energy; Then, extract frequency-domain features from each direction of the three-axis acceleration signals, a total of 11 types of feature values, namely average frequency-domain amplitude, center frequency, mean square frequency, frequency variance, frequency amplitude variance, average frequency, median frequency, total power, average power, frequency corresponding to the maximum power, and ratio of low-frequency power to high-frequency power; Next, use wavelet transform to perform wavelet packet decomposition and wavelet packet reconstruction on each direction of the three-axis acceleration signals to obtain 8 wavelet energy ratios, 8 wavelet energy entropies, 1 wavelet energy spectrum entropy, and at the same time perform singular value decomposition on the three-axis acceleration signals to obtain 1 wavelet packet singular spectrum entropy; After completing the feature extraction in the time domain, waveform, frequency domain, and wavelet domain, obtain a vibration signal feature matrix from the three-axis acceleration signals; Finally, for the sound signal, perform feature extraction only in the time domain to obtain a sound signal feature matrix; Step 5: First, calculate the correlation coefficients between the time-domain features, frequency-domain features, and wavelet-domain features of the three-axis acceleration signals and the average value of the VB values of the three flank faces of the tool, and then calculate the correlation coefficient between the time-domain features of the sound signal and the average value of the VB values of the three flank faces of the tool; Step 6: Screen out the features with a correlation coefficient greater than 0.5 with the average value of the VB values of the three flank faces of the tool to obtain a screened feature matrix. Perform a mean-zero operation on each element in the screened feature matrix, and then realize the fusion of the three-axis acceleration signals, sound signal features, and the intrinsic features of the rotating tool; Step 7: Divide the wear levels of the rotating tool into 6 categories, as shown in the following table. Among them, the wear value in the tool wear value interval refers to the average value of the wear bandwidth of the three flank faces of the tool; Step 8: Establish a feature fusion graph data set for tool wear; Step 9: Use model fusion technology to establish an integrated neural network architecture and perform training and verification. Specifically: Build an integrated neural network based on the feature fusion graph, mainly composed of Network 1, Network 2, Network 3, and a voting component; Network 1 serves as the upper layer architecture of the integrated neural network, with a structure of 2 convolutional activation layers, 3 fully connected layers, and 1 Softmax classifier; Network 2 serves as the middle layer architecture of the integrated neural network, with a structure of a combined layer of 2 convolutional activation layers and 1 convolutional pooling layer from left to right, a combined layer of 2 convolutional activation layers and 1 convolutional pooling layer, a combined layer of 3 convolutional activation layers and 1 convolutional pooling layer, a combined layer of 3 convolutional activation layers and 1 convolutional pooling layer, a combined layer of 3 convolutional activation layers and 1 convolutional pooling layer, 2 fully connected activation layers with loss, 1 fully connected layer, and 1 Sotfmax classifier; Network 3 serves as the lower layer architecture of the integrated neural network, with a structure of a combined layer of 1 convolutional activation layer and 1 convolutional pooling layer, a combined layer of 1 convolutional activation layer and 1 convolutional pooling layer, a combined layer of 3 convolutional activation layers and 1 convolutional pooling layer, 2 fully connected activation layers with loss, 1 fully connected activation layer, 1 fully connected layer, and 1 Softmax classifier; Set the loss function and iterator, train the integrated neural network based on the feature fusion graph using the training set, establish the integrated neural network model metrics based on the feature fusion graph, and verify the integrated neural network based on the feature fusion graph on the test set.

2. The method for monitoring the wear of a rotary tool based on a feature fusion graph of acoustic and vibration signals according to claim 1, wherein: Step 3 is specifically as follows: The machine tool drives the rotary tool to machine the workpiece according to the set cutting parameters and feed path. During the machining process, once the signal curves of the triaxial acceleration sensor and the sound level meter show abnormalities, or after one feed is completed, the spindle of the machine tool stops rotating and rises for resetting, the triaxial acceleration sensor and the sound level meter stop signal acquisition, the working platform moves to align the image acquisition device with the rotary tool, the horizontal industrial camera takes images of the flank wear area of the side cutting edge of the rotary tool, measures the VB values corresponding to the three cutting edges, and the vertical industrial camera takes images of the wear of the bottom cutting edge of the rotary tool; Before the next feed, store the triaxial acceleration signal, sound signal, VB value, and wear images; Conduct tests on 3 rotary tools, and each rotary tool records the triaxial acceleration signal, sound signal, VB value, and wear images corresponding to n = 315 feeds.

3. The method for monitoring the wear of a rotary tool based on a feature fusion graph of acoustic and vibration signals according to claim 1, wherein: Step 6 is specifically as follows: Step1. Feature screening is performed on the time domain, frequency domain, and wavelet domain features of the acceleration signals in each direction of the three-axis acceleration signal. Six features with a correlation coefficient greater than 0.5 with the average value of the VB values of the three flank faces of the cutting edges are selected and recorded in an n×6 screened feature matrix, where n is the number of tool passes and 6 is the number of features after screening. Then, the elements in the screened feature matrix are mean-centered to 0. Similarly, feature screening is performed on the time domain features of the acoustic signal. Six features with a correlation coefficient greater than 0.5 with the average value of the VB values of the three flank faces of the cutting edges are selected and recorded in an n×6 screened feature matrix, and the elements in this screened feature matrix are also mean-centered to 0. Step2. Calculate the eigenvalue covariance matrix corresponding to each element in the screened feature matrix obtained from the acceleration signals in each direction after mean-centering to 0. Then, calculate the eigenvalues and eigenvectors of each covariance matrix. Sort the distinct eigenvalues of each covariance matrix from largest to smallest. Set the number of principal components to 3. Analyze the principal component variance contribution rate after processing the eigenvectors corresponding to the eigenvalue solutions through PCA. Take the elements in the three eigenvectors with the largest cumulative variance contribution rate as the principal features of the corresponding direction acceleration signal. Step3. Sort the eigenvalues corresponding to each element in the screened feature matrix obtained from the acoustic signal after mean-centering to 0 from largest to smallest, and take the first two features as the principal features of the acoustic signal. Step4. Unitize the cutting time according to the number of tool passes. The time corresponding to one tool pass is defined as one cutting unit. Finally, normalize the number of tool passes to obtain the intrinsic features of the rotating tool. Step5. Combine the principal features of the acceleration signals in each direction, the principal features of the acoustic signal, and the intrinsic features of the rotating tool to obtain the fusion features.

4. The method for monitoring the wear of a rotating tool based on the feature fusion diagram of acoustic and vibration signals according to claim 3, characterized in that: Step eight is specifically as follows: Use circular polar coordinates as the coordinate system for the fusion features. Based on the 12 features of the fusion features, 12 polar axes are equally spaced around the origin. The values on each polar axis are restricted between 0 and 1. Take the number of tool passes as the first axis, and the 2nd to 10th axes are sequentially assigned to the 9 principal features of the three-axis acceleration signal. The remaining 2 axes are assigned to the principal features of the acoustic signal. For each tool pass, mark the values of the fusion features on their respective feature polar axes in sequence. Connect these points in sequence to form a closed loop. After establishing the feature fusion diagrams for all tool passes, establish a training set and a test set with a ratio of 8:

2.

5. The method for monitoring the wear of a rotating tool based on the feature fusion diagram of acoustic and vibration signals according to claim 1, characterized in that: In step nine, the indicators for establishing the integrated neural network model based on the feature fusion diagram are as follows: The first-level indicators for constructing an integrated neural network based on a feature fusion graph are TP, FN, FP, and TN. Among them, TP represents the number of positive samples where the true value is positive and the integrated neural network also considers it positive; FN represents the number of positive samples where the true value is positive but the integrated neural network considers it negative; FP represents the number of negative samples where the true value is negative but the integrated neural network considers it positive; TN represents the number of negative samples where the true value is negative and the integrated neural network also considers it negative.

6. A method for monitoring the wear of a rotary tool based on a feature fusion graph of acoustic and vibration signals according to claim 5, wherein: The second-level indicators for constructing an integrated neural network based on a feature fusion graph are Accuracy, Precision, Recall, and Specificity.

Images