Cutter health state prediction method based on fusion information

Abstract

The invention discloses a tool health state prediction method based on fusion information, and the method comprises the steps: collecting a tool wear image, main shaft vibration information and power information in a machining process of a machine tool, and the total use time of a tool from brand new to failure, and obtaining the remaining service life of the tool at any time; preprocessing the tool wear image, the spindle vibration information and the power information to obtain four target features, namely a tool wear area image, a VB value, a sensitive feature of spindle vibration to the remaining service life of the tool and a sensitive feature of machine tool power to the remaining service life of the tool; the target features are used as features for prediction, the remaining service life of the tool is used as a prediction target, and prediction of the remaining service life of the tool is trained; and processing the test set to obtain four target features, and inputting the four target features into the trained residual service life prediction of the cutter to obtain a residual service life prediction value of the cutter. The tool wear state can be monitored comprehensively and visually, and high-instantaneity and high-accuracy prediction is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of tool health status prediction, and in particular to a tool health status prediction method based on fusion information. Background Art

[0002] The tool is an important part of the machining process. As the cutting process progresses, the tool will continue to wear, and eventually the quality of the machining can no longer be guaranteed or even the machining can no longer be continued. At the same time, the wear of the tool will continue to increase the machining error, affecting the shape and size of the workpiece, and thus affecting the quality of the processed product. Studies have shown that tool health prediction technology with sufficient accuracy and robustness can reduce the downtime caused by technical and human subjective factors to about 25% of the original, while increasing production efficiency by at least 10%.

[0003] Therefore, it is particularly important to predict the health of the tool. On the one hand, knowing the wear of the tool can be used to make appropriate compensation during the processing to reduce errors; on the other hand, knowing the health of the tool can maximize the tool's service life without failure due to tool failure, so as to reduce the manufacturing costs caused by tool change downtime and the tool itself. However, the tool wear process is unstable, and its degradation trend is difficult to predict using traditional methods. Summary of the invention

[0004] The purpose of the present invention is to overcome the shortcomings and defects of the prior art and to provide a tool health status prediction method based on fusion information, which is a tool health status prediction method that integrates image features, spindle vibration features and power features during machine tool processing.

[0005] A tool health status prediction method based on fusion information includes the following steps:

[0006] Collect tool wear images, spindle vibration information, and machine power information during machine tool processing, which are used to train the tool remaining service life prediction model, as well as the total usage time of the tool from brand new to failure; the remaining service life of the tool at any time is obtained through the total tool usage time;

[0007] The tool wear image, spindle vibration information and power information are preprocessed to obtain the normalized tool wear area image, the wear width VB value of the tool side edge, the sensitive characteristics of the spindle vibration to the remaining service life of the tool, and the sensitive characteristics of the machine tool power to the remaining service life of the tool, a total of four target features;

[0008] Taking the target feature as the feature used for prediction, taking the remaining service life of the tool as the prediction target, using the training set data to train the constructed remaining service life prediction model of the tool, and using the validation set to verify;

[0009] After processing the test set data, the four target features are obtained and input into the trained tool remaining service life prediction model to obtain the remaining service life prediction value of the tool.

[0010] The wear width VB value of the side edge of the tool is obtained by the following steps:

[0011] The original image of the side edge of the tool is converted from the RGB color domain to the HSV color domain, and the initial binarization threshold segmentation is performed in the HSV color domain to extract the contour image of the tool cutting edge;

[0012] The tool cutting edge contour image is subjected to straight line recognition to extract the position of the tool cutting edge and the slope of the cutting edge. Then, secondary threshold segmentation is performed to obtain the side edge wear area image and convert it into a grayscale image.

[0013] Rotate the side edge wear area image in the form of a grayscale image so that the cutting edge is vertical, then calibrate the pixels to find the wear width VB value of the side edge;

[0014] The wear width VB value data of the side cutting edge is normalized to obtain the normalized wear width VB value of the tool side cutting edge.

[0015] Among them, when rotating the side blade wear area image in the form of a grayscale image, first, based on the obtained blade slope, the relationship between the image's inclination angle and the image's expected counterclockwise angle is determined, and then the image is rotated based on the relationship so that the cutting edge in the image is in a vertical state; when rotating, first rotate the image 45 degrees clockwise, and then use its diagonal as the length and height of the new image, and fill the missing data of the new image with a value of 0, i.e., black; then rotate the new image counterclockwise by θ+45 degrees, where θ is the new image's expected counterclockwise angle.

[0016] The wear width VB of the side edge is calculated by the following formula:

[0017] VB=max(b1,b2,b3,···,b n )×μ

[0018]

[0019] In the formula, are the maximum and minimum x-coordinates of all points in the i-th row of the new image after rotation, and b i is the pixel width of the i-th row in the wear area, n is the height of the new image, and μ is the scale of the new image, that is, the actual width of 1 pixel.

[0020] The wear width VB value of the tool side edge is normalized based on the full data, and the expression is as follows:

[0021]

[0022] Among them, VB MAx is the maximum value among all VBs. MIN is the minimum value among all VBs, VB is The normalized wear width VB of the tool side edge is represented by VB i Indicates the wear width VB value of the tool side edge before normalization.

[0023] The acquisition of the sensitive characteristics of the spindle vibration to the remaining service life of the tool and the sensitive characteristics of the machine tool power to the remaining service life of the tool includes:

[0024] After denoising the originally collected spindle vibration and power data, the characteristic values ​​of the vibration and power signals are obtained, and then the correlation analysis between the characteristic values ​​and the remaining service life of the tool is performed to obtain the correlation coefficient. A dimensionless index is constructed based on the correlation coefficient, and based on the dimensionless index, features that are insensitive to the remaining service life of the tool are excluded, and sensitive features that are sensitive to the remaining tool life are extracted. After normalization processing, the sensitive features of the spindle vibration to the remaining service life of the tool and the sensitive features of the machine tool power to the remaining service life of the tool are obtained.

[0025] The expression for constructing the dimensionless index based on the correlation coefficient is as follows:

[0026]

[0027] range=r max -r min

[0028]

[0029] Among them, coef represents a dimensionless index, mean and std represent the mean and standard deviation of the corresponding correlation coefficient of each tool under each eigenvalue, r max With r min They represent the maximum and minimum values ​​of the correlation coefficient r corresponding to each tool under each eigenvalue.

[0030] The sensitive characteristics of the spindle vibration to the remaining service life of the tool include the mean value mv, root mean square rms, frequency domain feature average power spectrum mps, standard deviation std and waveform factor sf of the vibration signal;

[0031] The sensitive characteristics of the machine tool power to the remaining service life of the tool include the root mean square rms of the minimum power, the mean value mv, the standard deviation std, the frequency domain characteristic average power spectrum mps and the waveform factor sf.

[0032] The tool remaining service life prediction model includes a convolutional neural network (CNN) for extracting tool image features and three BiLSTM models for learning the sensitive features of the wear width VB value of the tool side edge, the vibration signal of the spindle and the power signal to the tool service life; an artificial neural network for outputting the tool service life prediction value after processing the fusion features output by the three BiLSTM models and the convolutional neural network (CNN); each BiLSTM model is composed of a double-layer BiLSTM network stack, each BiLSTM network has the same sequential structure and different hyperparameters; the hyperparameters adopt the Hyperband algorithm, which is based on the Successive Halving algorithm to introduce an elimination system into the grid search method, and the optimal hyperparameter combination is obtained by elimination, including:

[0033] Allocate equal resources to n models of the same architecture with different hyperparameters, train in round A under the same conditions, eliminate a certain number of models, and further train in round B, eliminate a certain number of models, until the number of remaining models is 1, and obtain the optimal hyperparameter combination.

[0034] Among them, when the tool remaining service life prediction model processes data, the real-time collected tool wear area image, VB value, vibration and power signal, which are sensitive features of the tool remaining service life, are processed by the convolutional neural network CNN and three 2*BiLSTM networks respectively, and then nonlinearly fused into several abstract fusion features. Each abstract fusion feature is composed of the wear area image, VB value, vibration and power signal, and then the artificial neural network judges the tool remaining life based on the fusion features.

[0035] The present invention takes the remaining service life of the tool as the prediction object, is based on the improved CNN-BiLSTM integrated network as the backbone, takes the image features of the tool wear area and the wear width VB value of the tool side edge obtained based on the image features of the tool wear area, and the vibration and power signal characteristics of the spindle in machine tool processing as the prediction basis, can comprehensively and intuitively monitor the tool wear status, achieve high immediacy and high accuracy prediction, and has good prediction ability in the early stage of tool wear, with an average prediction error of 4.76min and a prediction time of less than 1s. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 The figure is a flow chart of an implementation method of a tool health status prediction method based on fusion information according to an embodiment of the present invention.

[0037] Figure 2 This is a processing path diagram for collecting raw data of tool wear in an embodiment of the present invention.

[0038] Figure 3 This is a flow chart of wear data collection for a tool throughout its life cycle according to an embodiment of the present invention.

[0039] Figure 4 This is the original image of the side edge of the tool according to the embodiment of the present invention.

[0040] Figure 5 This is a side edge profile diagram of a tool according to an embodiment of the present invention.

[0041] Figure 6 This is a straight line position diagram of the side edge of the tool according to an embodiment of the present invention.

[0042] Figure 7 This is a grayscale image of the side edge wear area of ​​the tool according to an embodiment of the present invention.

[0043] Figure 8 This is a structural diagram of a CNN model for extracting tool wear image features according to an embodiment of the present invention.

[0044] Fig. 9 This is a structural diagram of the BiLSTM model for learning tool wear VB, vibration and power sensitivity characteristics in an embodiment of the present invention.

[0045] Fig.10 This is the overall block diagram of the CNN-BiLSTM model based on fusion data in the present invention.

[0046] Fig.11 The figure is a logic block diagram of the Hyperband algorithm for selecting hyperparameters of the CNN-BiLSTM model based on fusion data in the present invention.

[0047] Fig.12 This is a diagram of the prediction effect of the CNN-BiLSTM model based on fused data in the present invention. DETAILED DESCRIPTION

[0048] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0049] With the development of computer technology, chip manufacturing technology and other technologies, artificial intelligence networks have brought a new possibility for tool wear stage monitoring and remaining service life prediction. The universal approximation theorem states that a neural network with at least one hidden layer can fit an arbitrary function with arbitrary precision. At the same time, compared with the previous artificial features, the use of neural network self-extracted features can more accurately and comprehensively reflect the wear of the tool. Therefore, a properly trained neural network is fully expected to surpass the traditional research content in the past, more accurately fit the wear degradation process of the tool, and make more accurate and reliable predictions of the wear condition and remaining useful life (RUL) of the tool.

[0050] To this end, the inventors of the present application have, through creative work, proposed a method for predicting tool health status based on fused information of the present invention, in particular, a method for multi-source data fusion and tool remaining service life prediction based on an improved CNN-BiLSTM model, which utilizes multi-source information including images, vibrations, power, etc. collected during the milling cutter processing to accurately predict the remaining service life of the tool, thereby guiding tool sharpening and replacement, etc., to achieve the purpose of reducing manufacturing costs, improving processing efficiency, and ensuring processing accuracy.

[0051] In order to realize the present invention, the embodiment of the present invention first designs a collection system for characteristic data related to the remaining service life of the tool to perform relevant data collection, and then establishes a wear database for the tool throughout its life cycle based on the data collection of image, vibration and power information; then, the collected raw data is preprocessed to remove redundant information in multi-source information, and the wear area image of the side blade of the milling cutter is extracted based on the method of straight line recognition and double threshold segmentation, thereby achieving a basically complete extraction of the side blade wear area; then, based on the correlation analysis method, the sensitive characteristics of vibration and power signals to the remaining service life of the tool are determined, and whether the tool has failed is judged based on the determined tool failure standard.

[0052] In the embodiment of the present application, the tool failure index is determined based on characteristic data related to the remaining service life of the tool. The tool can be an end mill or a similar tool of a high-speed and high-precision equipment that operates normally in an intelligent factory.

[0053] In the embodiment of the present application, preprocessing of the collected raw data is a method for improving the model fitting accuracy and training speed, and the preprocessing includes preprocessing of the collected tool wear area image, the vibration of the tool spindle and the power signal.

[0054] In an embodiment of the present application, a tool remaining service life prediction model is constructed based on an improved CNN-BiLSTM. The model conducts deep learning on the feature information of preprocessed data, fuses data from different sources using an artificial neural network, and finally predicts the tool remaining service life based on the fused information.

[0055] like Figure 1 As shown, in an embodiment of the present application, a tool health status prediction method based on fusion information includes the steps of:

[0056] S1: First, collect the tool wear image, spindle vibration information and power information during the machine tool processing, which are necessary for the tool remaining service life prediction model, and the total use time of the tool from brand new to failure; through the total use time of the tool, the remaining service life of the tool at any time is obtained;

[0057] S2: For tool wear images, spindle vibration information and power information, use appropriate data preprocessing methods to eliminate useless information and enhance effective information. Through preprocessing of tool wear images, spindle vibration information and power information, obtain the normalized tool wear area image, tool side edge wear width VB value, spindle vibration sensitivity to tool remaining service life, power sensitivity to tool remaining service life, a total of four target features;

[0058] S3: The target features described in the four items and the tool remaining service life obtained in the previous experiment are used as input data, the tool wear area image after preprocessing, the wear width VB value of the tool side edge, the spindle vibration sensitivity to the tool remaining service life and the power sensitivity to the tool remaining service life are used as prediction features, the tool remaining service life is used as the prediction target, and the constructed tool remaining service life prediction model is trained and verified with the verification set;

[0059] After the training of the tool remaining service life prediction model is completed, based on the four pre-processed feature data, the tool remaining service life prediction value with sufficient accuracy is output;

[0060] S4: When predicting the tool, the four target features obtained after processing the data of the test set are input into the trained tool remaining service life prediction model to obtain the remaining service life prediction value of the tool.

[0061] In the embodiment of the present application, the remaining service life prediction model of the tool obtained is the RUL (remaining tool life) prediction model:

[0062] Tool wear is generally divided into three stages, namely initial wear, normal wear, and rapid wear. When the tool enters the rapid wear stage, the contact area between the back of the tool and the machining surface is larger, and the cutting temperature and cutting force increase rapidly. Therefore, the instability in the machining process increases sharply, while the machining accuracy and production efficiency decrease rapidly. In production, in order to ensure the machining quality, the tool should be changed before rapid wear occurs. Tool wear mainly occurs on the side edge, that is, the back face of the tool during side milling. Therefore, in this application, the collection and subsequent prediction of tool data mainly focus on the side edge wear of the tool (end mill).

[0063] In the experiment, the lean four-axis machining center of Tianjin University was used, and the HIOKIPW3360 clamp power meter was used to measure the machine power signal. The average, minimum and maximum power between two archives were saved once per second. A piezoelectric sensor was installed in the y direction of the machine tool spindle to collect the vibration signal of the machine tool spindle at 2000Hz; a CCD industrial camera was used to collect the side edge image of the end mill, and its parameters are shown in Table 1 below. A ring-shaped adjustable light source was used to provide lighting to ensure stable lighting and reduce the difficulty of subsequent processing. After calibration, the width of each pixel of the captured image is 2.082031μm.

[0064] Table 1

[0065]

[0066] Among them, there is a controllable magnetic device at the bottom of the CCD industrial camera. It is installed at a predetermined position when taking pictures (non-processing stage). After installation, the workbench can be manually moved to a fixed shooting position to take pictures of the tool. Since the position of the blade after processing is uncontrollable, the four cutting edges of the end mill are manually marked on the handle with a marker pen to distinguish them. At the same time, by marking on the display screen to ensure the approximate position of the blade wear area, the difficulty of post-processing is reduced. The vibration sensor is adsorbed on the spindle to collect active vibration signals.

[0067] The experimental data is based on a 10mm diameter white steel end mill, which is repeatedly milled on 60mm×60mm×40mm 45 steel. The image of the milling cutter side edge is collected every 6 passes. The processing path of the 6 passes is as follows: Figure 2 As shown in the figure, the white steel end mill moves in a serpentine path on the 45 steel. The VB value is manually measured using the S-EYE software supporting the CCD industrial camera. The cutting parameters are shown in Table 2.

[0068] Table 2

[0069]

[0070] In the experiment, the process of collecting wear information of each tool Figure 3 As shown. First, the vibration and power signal collection starts at the beginning of processing. After completing 6 tool passes according to the path, the vibration and power signal collection is paused and the current time point is marked. Then, the side edge image of each cutting edge is captured by the camera system, and the VB value of each cutting edge is manually measured using software. If the VB value of a cutting edge is greater than 300μm, the tool is considered to be ineffective and the experiment is stopped. Otherwise, repeat the above steps.

[0071] The above are the steps and methods for collecting data of images, vibration and power signals in the embodiment of the present application. The above is only an embodiment, and the collection equipment or method used in the specific implementation of the present invention may not be limited to the equipment and method mentioned above, as long as the spindle vibration signal and power signal and tool wear image required by the present invention can be collected.

[0072] In the embodiment of the present application, the original image of the side edge of the end mill taken in the experimental environment is as follows: Figure 4 As shown. Many contents in the original image of the side blade, including the background, are redundant information, which is not helpful in evaluating the wear condition of the tool or predicting the remaining service life of the tool. Instead, they may become interference information that affects the prediction process. Therefore, in the embodiment of the present application, after the tool image is collected, the required feature extraction is performed on the wear area of ​​the tool using an available processing method or technology to obtain the required wear area image of the side blade, and the wear width VB value of the side blade is calculated and normalized. The implementation steps include:

[0073] S21. Converting the initial image or original image of the side edge of the tool from the RGB color domain to the HSV color domain, performing initial binarization threshold segmentation in the HSV color domain, and extracting the tool cutting edge contour image;

[0074] The conversion formula from RGB color gamut to HSV color gamut is as follows:

[0075] V=max(R,G,B)

[0076]

[0077] In the formula, R, G, and B correspond to the three-channel color values ​​in the RGB color domain, and H, S, and V correspond to the three-channel color values ​​in the HSV color domain. By setting a suitable threshold in the HSV color domain, the tool cutting edge contour image can be preliminarily extracted, such as Figure 5 As shown, Figure 5 This is the preliminary extracted tool cutting edge contour image.

[0078] S22. Figure 5 Based on the tool cutting edge contour image shown, straight line recognition is performed to extract the position of the tool cutting edge and the blade slope k0, and then secondary threshold segmentation is performed to obtain the side edge wear area image and convert it into a grayscale image. Optionally, the straight line recognition method used is as follows:

[0079] In the image space xy, a straight line can be expressed in the rectangular coordinate system as: y = kx + b, where k and b are parameters, representing the slope and intercept of the line. The parameters of all straight lines passing through a certain point A (x_0, y_0) satisfy the equation y_0 = kx_0 + b, that is, point A (x_0, y_0) determines a family of straight lines. Rewrite the equation as: b = -x_0k + y_0, then the family of straight lines determined by point A (x_0, y_0) corresponds to a straight line in the parameter space kb. The straight line determined by point A (x_0, y_0) and point B (x_1, y_1) corresponds to a unique point in the parameter space kb, that is, the intersection of the family of straight lines determined by point A in the parameter space kb corresponding to the straight line l_A and the family of straight lines determined by point B in the parameter space kb corresponding to the straight line l_B.

[0080] Therefore, we only need to find that all points X_i on the image correspond to lines l_i in the parameter space kb, and the point with the most intersections of lines (k_0, b_0) can determine the line y=k_0x+b_0 where the blade is located. Figure 5 The straight line required is Figure 6 As shown by the red line in the middle, the straight line is translated left and right, and only the sandwiched area is retained to obtain the blade position. Then, secondary threshold segmentation is performed in the HSV color domain, and a suitable threshold is selected to obtain the side blade wear area image. On this basis, the image is converted into a grayscale image to reduce the amount of data, as shown in Figure 7 shown.

[0081] S23. Rotate the side edge wear area image to make the cutting edge vertical, then calibrate the pixels to find the side edge wear width VB value.

[0082] The blade slope k0 has been obtained in the above steps, and the inclination angle θ0 and the image counterclockwise angle θ are obtained by the following formula:

[0083]

[0084] θ=90-θ0

[0085] Based on the relationship between the inclination angle θ0 and the counterclockwise angle θ of the image, the image of the side edge wear area is rotated to make the cutting edge vertical, and then the pixels are calibrated to calculate the wear width VB value of the side edge.

[0086] When rotating an image, first rotate the image 45 degrees clockwise, use its diagonal as the length and height of the new image, and fill the missing data of the new image with 0 value, i.e. black; then rotate the new image counterclockwise by θ+45 degrees to avoid local information loss.

[0087] Among them, the cutting edge is in a vertical state, the pixels are calibrated, and the wear width VB value of the side edge is calculated by the following formula:

[0088]

[0089] VB=max(b1,b2,b3,···,b n )×μ

[0090] In the formula, are the maximum and minimum x-coordinates of all points in the i-th row of the new image, respectively, and b i is the pixel width of the i-th row in the wear area, n is the height of the new image, and μ is the scale of the new image, that is, the actual width of 1 pixel.

[0091] S24. Normalize the image of the cutting edge vertically obtained in step S23 and the wear width VB value data of the side edge. The normalization process facilitates the subsequent training of the tool remaining service life prediction model.

[0092] Among them, the grayscale image of the side edge wear area can be expressed as:

[0093] GWI r =[p ij ] 1080×1920

[0094] In the formula, p ik Indicates the gray value of each point in the side edge wear area. r Represents the grayscale image of the side edge wear area.

[0095] Since the grayscale image data is a two-dimensional matrix with a minimum value of 0 and a maximum value of 255, the normalization process of the grayscale image of the side edge wear area can be simply expressed as:

[0096]

[0097] In the formula, GWI s Represents the normalized grayscale image of the side edge wear area.

[0098] The wear width VB value of the tool side edge should be normalized based on the full data, that is:

[0099]

[0100] Among them, VB MAX is the maximum value among all VBs obtained based on the data. MIN is the minimum value of all VBs obtained from the data. is The normalized wear width VB of the tool side edge is represented by VB i Represents the wear width VB value at different positions of the tool side edge before normalization.

[0101] The various features in the time domain can directly reflect the instantaneous changes of the signal, among which the mean value MV, the root mean square RMS, and the standard deviation STD reflect the amplitude and central tendency of the signal, and the waveform factor SF, the skewness SKE, and the kurtosis KUR reflect the waveform of the signal. The wear of the tool is reflected in the vibration of the spindle and the frequency characteristics of the power signal at the same time. Therefore, the embodiment of the present application selects three typical frequency domain features, the average power spectrum MPS, the frequency centroid FC, and the mean square frequency MSF, which respectively reflect its power spectrum, centroid position, and main frequency band.

[0102] Therefore, the processing of spindle vibration and power data related to tool wear in the embodiment of the present application includes:

[0103] After denoising the original collected spindle vibration and power data, the characteristic values ​​of the vibration and power signals are obtained, and then the correlation analysis of the remaining service life of the tool is carried out. The features insensitive to the remaining service life are excluded, and the sensitive features sensitive to the remaining service life are extracted. After normalization, the sensitive features of the spindle vibration to the remaining service life of the tool and the sensitive features of the machine tool power to the remaining service life of the tool are obtained.

[0104] Since not all the characteristic values ​​are affected by the tool usage time or the remaining tool life, the correlation analysis between the characteristic information of the spindle vibration and power signal and the remaining tool life is performed to exclude the characteristics that are insensitive to the remaining service life; specifically, this can be achieved through the following steps:

[0105] S31. Calculate the correlation coefficient between each characteristic value and tool life:

[0106] The above correlation coefficient can be expressed as follows:

[0107]

[0108] Where r represents the correlation coefficient between each eigenvalue and tool life, and x and y represent different types of eigenvalues.

[0109] S32. Construct a dimensionless index for the correlation coefficient between each characteristic value and the tool life.

[0110] Since each signal involves the data of multiple tools, and there are differences between the data of each tool, the performance of each feature cannot be quantitatively measured by the obtained correlation coefficient alone. Therefore, for the correlation coefficient between each eigenvalue and the tool life, a dimensionless index is constructed. The dimensionless index is used to measure the overall correlation of each eigenvalue as the standard for selecting sensitive features; the expression of the constructed dimensionless index is as follows:

[0111]

[0112] range=rmax -r min

[0113]

[0114] Among them, coef represents a dimensionless index, mean and std represent the mean and standard deviation of the corresponding correlation coefficient of each tool under each eigenvalue, r max With r min They represent the maximum and minimum values ​​of the correlation coefficient r corresponding to each tool under each eigenvalue.

[0115] S33. Based on the dimensionless index coef, the most sensitive features are extracted for subsequent model training and tool remaining service life prediction.

[0116] Specifically, among the various power signals, the various features of the minimum power value have the best correlation with the remaining service life of the tool, among which the root mean square rms, mean mv, standard deviation std, frequency domain feature average power spectrum mps, and waveform factor sf are the best; among the various features of the vibration signal, the mean mv, root mean square rms, mps, standard deviation std, and waveform factor sf have the best correlation. Therefore, the root mean square rms of the minimum power value, the mean mv, standard deviation std, the frequency domain feature average power spectrum mps, and waveform factor sf are constructed as the five sensitive features of the power signal; the mean mv, root mean square rms, frequency domain feature average power spectrum mps, standard deviation std, and waveform factor sf of the vibration signal are constructed as the five sensitive features of the vibration signal. The correlation index coef of the vibration value, maximum power, minimum power, and average power is shown in Table 3 below.

[0117] Table 3

[0118]

[0119] S34. Normalize each sensitive feature within its scope based on the entire data to solve the problem of inconsistent magnitude and dimension of different sensitive features:

[0120]

[0121] Among them, X MAX is the maximum value of all sensitive features obtained from the data, X MIN is the minimum value of all sensitive features obtained from the data, X i The data representing the sensitive features to be processed, X is Represents the data of sensitive features after normalization.

[0122] In an embodiment of the present application, the tool remaining service life prediction model is a tool health status prediction model based on fused data, which learns tool wear image features based on a convolutional neural network (CNN) model.

[0123] Convolutional neural network is a neural network that recognizes and predicts images. Compared with the classic feedforward neural network, CNN significantly reduces the number of network parameters, improves learning and reasoning speed, and has translation invariance. It can also use methods such as image enhancement to improve its rotation and scaling invariance. It is generally composed of convolutional layers, pooling layers, and fully connected layers. The basic network structure of CNN is Figure 8 shown.

[0124] The forward propagation algorithm of the convolutional layer is as follows:

[0125]

[0126] X (l) =a(Z (l) )

[0127] Where, X (l-1) is the input tensor, W (l) is the convolution kernel tensor, b (l) is the bias tensor, Z (l) is the net output tensor, a(·) is the activation function, X (l) is the output tensor, symbol Represents the convolution operation. The activation function a(·) and the convolution kernel tensor W (l) The shape of W needs to be given in advance as a hyperparameter. (l) With b (l) The value of is learned by back propagation method. The CNN network structure used for extracting tool image features of the present invention is listed in Table 4 below, and its final output is a sequence of 7 features:

[0128] Table 4

[0129]

[0130] As shown in the table above, the CNN network structure for extracting tool image features sequentially inputs the input layer Input, the convolution layer Conv (the activation function uses Relu), the maximum pooling layer Maxpooling, the convolution layer Conv (the activation function uses Relu), the average pooling layer Averagepooling, the flattening layer Flatten, the temporary retreat layer Dropout, and the fully connected layer Dense (the activation function uses Leaky Relu). The fully connected layer Dense is implemented using an artificial neural network, and the tool image features x at different positions output by the previous layer are converted into nAfter dimensionality reduction processing, the dimensionality reduction feature h is output n , and finally the target features of the tool wear area image are obtained after learning.

[0131] In the present application, the tool remaining service life prediction model is based on the BiLSTM model to learn the sensitive characteristics of the VB value, the vibration signal of the spindle and the power signal. Among them, for the sensitive characteristics of the VB value, the vibration signal of the spindle and the power signal, the present invention uses a bidirectional long short-term memory (Bidirectional Long Short-Term Memory, BiLSTM) neural network for processing. LSTM network is a type of recurrent neural network (Recurrent Neural Network, RNN). Compared with S-RNN, the LSTM network introduces memory elements and gating mechanisms, has a stronger ability to handle long-distance dependencies, and effectively avoids the gradient vanishing and gradient explosion problems of S-RNN. The BiLSTM network is connected using a bidirectional LSTM network, which can consider contextual information more comprehensively. Its basic network structure is as follows Fig. 9 As shown in the figure, three input values ​​X1, X2, and X3 are input into the network and are output after being processed by a two-layer bidirectional BiLSTM network.

[0132] The embodiment of the present invention uses a two-layer bidirectional BiLSTM network stack combined with a Reshape layer to mine deeper information. The network structure for processing the sensitive features of VB value, vibration signal and power signal is shown in Table 5 below.

[0133] Table 5

[0134]

[0135] As shown in the table above, the two-layer bidirectional BiLSTM network includes an input layer, a Reshape layer, a BiLSTM layer, a Reshape layer, and a BiLSTM layer in sequence, and processes the sensitive features of the VB value, the vibration signal, and the power signal, wherein the input VB value is a feature sequence, and the vibration signal and the power signal are five feature sequences. Through the bidirectional long and short time neural network layer by layer deep learning, the final output VB value is 10 feature sequences, the vibration signal is 34 feature sequences, and the power signal is 30 feature sequences.

[0136] Therefore, in the embodiment of the present application, the tool remaining service life prediction model has an overall network structure as follows: Fig.10As shown in FIG. 1 , the CNN-BiLSTM model includes a CNN network and a BiLSTM network. That is, for the wear area image of the tool, the CNN network is used to perform deep feature learning on the side edge wear area image; for the VB, vibration sensitive features to the remaining service life and power sensitive features to the remaining service life, a 2*BiLSTM network is used for deep feature learning. Each 2*BiLSTM network has the same sequential structure but different hyperparameters; during the forward reasoning process of the model, the wear area image, VB value, vibration and power signal sensitive features to the remaining service life of the tool will be transmitted to the head network through various parts of the backbone network, and will be nonlinearly fused into several abstract fusion features X in the head network. n , where each abstract fusion feature X n They are composed of the wear area image, VB value, vibration and power signal, and the dimension reduction feature H is formed through the Dense layer. n , and finally output the predicted life span Y.

[0137] In the forward reasoning process of the tool life model, the sensitive features of the wear area image, VB value, vibration and power signal to the remaining service life of the tool will be transmitted to the head network through various parts of the backbone network, and will be nonlinearly fused into several abstract fusion features in the head network. Each fusion feature is composed of the wear area image, VB value, vibration and power signal, but does not have a suitable physical explanation.

[0138] Feature fusion is learned by the model itself through the back propagation process of the prediction model: the error between the remaining tool life predicted by forward reasoning and the actual value will be transmitted to various parts of the model during the back propagation process, so that the model can automatically adjust its parameters and finally determine the composition of the fusion features.

[0139] For the tool remaining service life prediction model, its hyperparameters include the convolution kernel shape and activation function of each convolution layer; the kernel size of each pooling layer; the shielding rate of the dropout layer; the number of units and activation function of each fully connected layer (Dense); the number of units and sequence length of each BiLSTM layer; the training epoch; the batch size; the model learning rate, etc. In the process of optimizing the model hyperparameters, the Hyperband algorithm is used. This algorithm is based on the Successive Halving algorithm and introduces an elimination system to the grid search method. The optimal hyperparameter combination is obtained through elimination: that is, equal resources are allocated to n models of the same architecture with different hyperparameters, and a certain number of models are eliminated after training in round A under the same conditions. Then, a certain number of models are eliminated after further training in round B until the number of remaining models is 1, and the optimal hyperparameter combination is obtained. The program flow is as follows: Fig.11 shown.

[0140] In the embodiment of the present application, when training the tool remaining service life prediction model, the data is divided into a training set, a validation set, and a test set in a ratio of 0.64:0.16:0.2. The training set is used to train the model; the validation set is used to test the performance of the model after each round of training, and does not participate in the training itself; the test set is used to test the generalization prediction ability of the model after the model training is completed. The full-cycle prediction results of the 6 collected knives are as follows: Fig.12 shown.

[0141] It can be seen that the method model of the present invention maintains prediction results within ±6.8min throughout the entire process, has good early prediction capabilities, can obtain a relatively accurate remaining tool life in the early stages of tool wear, and is conducive to maximizing the use of tool performance.

[0142] The prediction effect of the tool remaining service life prediction model on image, VB, vibration and power. Under the same model structure, the performance of the model using only part of image, VB, vibration and power for model training and prediction is shown in Table 6 below.

[0143] Table 6

[0144]

[0145] Among the three indicators of image, vibration, and power, the prediction effect increases in order. Compared with using only vibration or power, the prediction effect after fusing vibration and power signals is 6.92 and 0.3 lower in MAE, 377.79 and 18.6 lower in MSE, and 15.94 and 2.2 lower in maximum absolute error. If image data is added on this basis, compared with fusing only vibration and power signals, MAE is reduced by 3.21 and MSE is reduced by 83.31. If the image, VB, vibration, and power information are fused, the optimal MAE, MSE, and maximum absolute error can be obtained, which are 4.76, 35.99, and 19.62, respectively. Its performance is better than any of the above combinations or using only VB.

[0146] Among them, MAE and MSE are the mean absolute error and mean square error respectively, and the calculation formula is as follows:

[0147]

[0148] In the formula, y i is the actual RUL of the i-th group of data, is the model's predicted RUL for the i-th group of data, and n is the total number of data groups in the test set.

[0149] The basic principles and main features of the present invention and the advantages of the present invention are shown and described above. It is obvious to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and the present invention can be implemented in other specific forms without departing from the spirit or basic features of the present invention.

[0150] Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is therefore intended that all changes falling within the meaning and range of equivalent elements of the claims are included in the present invention.

[0151] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.

Claims

1. A tool health status prediction method based on fusion information, characterized in that: Includes steps: Collect tool wear images, spindle vibration information, and machine power information during machine tool processing, which are used to train the tool remaining service life prediction model, as well as the total usage time of the tool from brand new to failure; the remaining service life of the tool at any time is obtained through the total tool usage time; The tool wear image, spindle vibration information and power information are preprocessed to obtain the normalized tool wear area image and the wear width VB value of the tool side edge, the sensitive characteristics of the spindle vibration to the remaining service life of the tool, and the sensitive characteristics of the machine tool power to the remaining service life of the tool, a total of four target features; Taking the target feature as the feature used for prediction, taking the remaining service life of the tool as the prediction target, using the training set data to train the constructed remaining service life prediction model of the tool, and using the validation set to verify; After processing the test set data, the four target features are obtained and input into the trained tool remaining service life prediction model to obtain the remaining service life prediction value of the tool.

2. The tool health status prediction method based on fusion information according to claim 1 is characterized in that: The wear width VB value of the tool side edge is obtained by the following steps: The original image of the side edge of the tool is converted from the RGB color domain to the HSV color domain, and the initial binarization threshold segmentation is performed in the HSV color domain to extract the contour image of the tool cutting edge; The tool cutting edge contour image is subjected to straight line recognition to extract the position of the tool cutting edge and the slope of the cutting edge. Then, secondary threshold segmentation is performed to obtain the side edge wear area image and convert it into a grayscale image. Rotate the side edge wear area image in the form of a grayscale image so that the cutting edge is vertical, then calibrate the pixels to find the wear width VB value of the side edge; The wear width VB value data of the side cutting edge is normalized to obtain the normalized wear width VB value of the tool side cutting edge.

3. The tool health status prediction method based on fusion information according to claim 2 is characterized in that: When rotating the side blade wear area image in the form of a grayscale image, first determine the relationship between the image's inclination angle and the image's counterclockwise angle based on the obtained blade slope, and then rotate the image based on the relationship so that the cutting edge in the image is in a vertical state; when rotating, first rotate the image 45 degrees clockwise, then use its diagonal as the length and height of the new image, and fill the missing data of the new image with a value of 0, i.e., black; then rotate the new image counterclockwise by θ+45 degrees, where θ is the counterclockwise angle of the new image.

4. The tool health status prediction method based on fusion information according to claim 3 is characterized in that: The wear width VB of the side edge is calculated by the following formula: VB=max(b1,b2,b3,···,b n )×μ In the formula, are the maximum and minimum x-coordinates of all points in the i-th row of the new image after rotation, and b i is the pixel width of the i-th row in the wear area, n is the height of the new image, and μ is the scale of the new image, that is, the actual width of 1 pixel.

5. The tool health status prediction method based on fusion information according to claim 4 is characterized in that: The wear width VB value of the tool side edge is normalized based on the full data, and the expression is as follows: Among them, VB MAX is the maximum value among all VBs. MIN is the minimum value among all VBs, VB is The normalized wear width VB of the tool side edge is represented by VB i Indicates the wear width VB value of the tool side edge before normalization.

6. The tool health status prediction method based on fusion information according to claim 1 is characterized in that: The acquisition of the sensitive characteristics of the spindle vibration to the remaining service life of the tool and the sensitive characteristics of the machine tool power to the remaining service life of the tool includes: After denoising the originally collected spindle vibration and power data, the characteristic values ​​of the vibration and power signals are obtained, and then the correlation analysis between the characteristic values ​​and the remaining service life of the tool is performed to obtain the correlation coefficient. A dimensionless index is constructed based on the correlation coefficient, and based on the dimensionless index, features that are insensitive to the remaining service life of the tool are excluded, and sensitive features that are sensitive to the remaining tool life are extracted. After normalization processing, the sensitive features of the spindle vibration to the remaining service life of the tool and the sensitive features of the machine tool power to the remaining service life of the tool are obtained.

7. The tool health status prediction method based on fusion information according to claim 6 is characterized in that: The expression for constructing the dimensionless index based on the correlation coefficient is as follows: range=r max -r min Among them, coef represents a dimensionless index, mean and std represent the mean and standard deviation of the corresponding correlation coefficient of each tool under each eigenvalue, r max With r min They represent the maximum and minimum values ​​of the correlation coefficient r corresponding to each tool under each eigenvalue.

8. The tool health status prediction method based on fusion information according to claim 1 is characterized in that: The sensitive characteristics of the spindle vibration to the remaining service life of the tool include the mean value mv, root mean square rms, frequency domain characteristic average power spectrum mps, standard deviation std and waveform factor sf of the vibration signal; The sensitive characteristics of the machine tool power to the remaining service life of the tool include the root mean square rms of the minimum power, the mean value mv, the standard deviation std, the frequency domain characteristic average power spectrum mps and the waveform factor sf.

9. The tool health status prediction method based on fusion information according to claim 1 is characterized in that: The tool remaining service life prediction model includes a convolutional neural network (CNN) for extracting tool image features and three BiLSTM models for respectively learning the wear width VB value of the tool side edge, the vibration signal of the spindle and the sensitive features of the machine tool power signal to the tool service life, and an artificial neural network for outputting the tool service life prediction value after processing the fusion features output by the three BiLSTM models and the convolutional neural network (CNN); each BiLSTM model is composed of a double-layer BiLSTM network stack, each BiLSTM network has the same sequential structure and different hyperparameters; the hyperparameters adopt the Hyperband algorithm, which is based on the Successive Halving algorithm and introduces an elimination system into the grid search method, and obtains the optimal hyperparameter combination through elimination, including: Allocate equal resources to n models of the same architecture with different hyperparameters, train in round A under the same conditions, eliminate a certain number of models, and further train in round B, eliminate a certain number of models, until the number of remaining models is 1, and obtain the optimal hyperparameter combination.

10. The tool health status prediction method based on fusion information according to claim 1, characterized in that: When the tool remaining service life prediction model processes data, the real-time collected tool wear area image, VB value, vibration and power signal, which are sensitive features of the tool remaining service life, are processed by the convolutional neural network CNN and three 2*BiLSTM networks respectively, and then nonlinearly fused into several abstract fusion features. Each abstract fusion feature is composed of the wear area image, VB value, vibration and power signal. Then, the artificial neural network judges the tool remaining life based on the fusion features.

Images