Tool health status prediction method based on fused information

Abstract

This invention discloses a tool health status prediction method based on fused information, comprising: acquiring tool wear images, spindle vibration information, and power information during machine tool processing, and the total service time of the tool from new to failure, to determine the remaining tool life at any given time; preprocessing the tool wear images, spindle vibration information, and power information to obtain four target features: tool wear zone image, VB value, sensitivity features of spindle vibration to the remaining tool life, and sensitivity features of machine tool power to the remaining tool life; training the tool life prediction function using the target features as the prediction features and the remaining tool life as the prediction target; inputting the four target features obtained after processing the test set into the trained tool life prediction function to obtain the predicted tool life. This invention can comprehensively and intuitively monitor tool wear status, achieving highly real-time and highly accurate prediction.

Description

Technical Field

[0001] This invention relates to the field of tool health status prediction technology, and in particular to a tool health status prediction method based on fused information. Background Technology

[0002] Cutting tools are a crucial component in the machining process. As cutting progresses, tools continuously wear down, eventually failing to guarantee machining quality or even halting further processing. Simultaneously, tool wear increases machining errors, affecting the shape and dimensions of the workpiece and consequently impacting product quality. Research indicates that tool health prediction technologies with sufficient accuracy and robustness can reduce downtime caused by technical and human subjective factors to approximately 25%, while simultaneously increasing productivity by at least 10%.

[0003] Therefore, predicting the health of cutting tools is particularly important. On the one hand, understanding the wear condition of the tools allows for appropriate compensation during machining to reduce errors; on the other hand, understanding the health of the tools allows for maximizing tool life without causing malfunctions due to tool failure, thereby reducing downtime for tool changes and the manufacturing costs associated with the tools themselves. However, the wear process of cutting tools is unstable, and its degradation trend is difficult to predict using traditional methods. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings and defects of the prior art and provide a tool health status prediction method based on fused information. This method 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 fused information includes the following steps:

[0006] The machine tool wear images, spindle vibration information, and machine tool power information during the machining process are collected to train the tool remaining service life prediction model. The total service time of the tool from brand new to failure is also collected. Based on the total service time of the tool, the remaining service life of the tool at any time is obtained.

[0007] Preprocessing of tool wear images, spindle vibration information, and power information yields normalized tool wear area images, tool side edge wear width VB value, spindle vibration sensitivity to remaining tool service life, and machine tool power sensitivity to remaining tool service life, for a total of four target features.

[0008] Using the target features as the features used for prediction and the remaining tool life as the prediction target, a tool life prediction model is trained using training set data and then verified using a validation set.

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

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

[0011] The original image of the side edge of the tool is converted from the RGB color gamut to the HSV color gamut. Initial binarization thresholding is performed in the HSV color gamut to extract the cutting edge contour image of the tool.

[0012] Line recognition is performed on the contour image of the cutting edge of the tool to extract the location of the cutting edge and the slope of the cutting edge. Then, secondary thresholding is performed to obtain the side edge wear area image and convert it into a grayscale image.

[0013] The side edge wear area image is rotated in grayscale form to make the cutting edge vertical, and then the pixels are calibrated to calculate the wear width VB value of the side edge.

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

[0015] When rotating the grayscale image of the side edge wear area, the relationship between the image's tilt angle and the counterclockwise angle to be rotated is first determined based on the obtained blade slope. Then, the image is rotated based on this relationship so that the cutting edge in the image is in a vertical state. During rotation, the image is first rotated 45 degrees clockwise, and then its diagonal is used as the length and height of the new image. The missing data in the new image is filled with 0 values, i.e., black. Subsequently, the new image is rotated counterclockwise by θ+45 degrees, where θ is the counterclockwise angle to be rotated in the new image.

[0016] The wear width VB value of the side blade is calculated using the following formula:

[0017] VB = max(b1, b2, b3, ..., b n )×μ

[0018]

[0019] In the formula, Let b be the maximum and minimum x-coordinates of all points in the i-th row of the rotated new image. i Let be the pixel width of the i-th row of the worn area, n be the height of the new image, and μ be the scale of the new image, i.e., the actual width of 1 pixel.

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

[0021]

[0022] Among them, VB MAx It is the maximum value in all VB files. MIN It is the minimum value in all VB files. is VB represents the normalized wear width of the tool side edge. i This represents the wear width VB value of the tool side edge before normalization.

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

[0024] After denoising the original acquired spindle vibration and power data, the characteristic values ​​of the vibration and power signals are obtained. Then, the correlation between the characteristic values ​​and the remaining tool life is analyzed to obtain the correlation coefficient. Based on the correlation coefficient, a dimensionless index is constructed. Based on the dimensionless index, features that are not sensitive to the remaining tool life are excluded, and sensitive features that are sensitive to the remaining tool life are extracted. Then, the data are normalized to obtain the sensitivity features of spindle vibration to the remaining tool life and the sensitivity features of machine tool power to the remaining tool life.

[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] Where, coef represents a dimensionless index, mean and std represent the mean and standard deviation of the correlation coefficients for each tool under each characteristic value, respectively, and r max With r min These represent the maximum and minimum values ​​of the correlation coefficient r for each tool under each characteristic value.

[0030] The sensitivity characteristics of the spindle vibration to the remaining tool life include the mean mv, root mean square rms, average power spectrum mps, standard deviation std, and waveform factor sf of the vibration signal.

[0031] The sensitivity characteristics of machine tool power to the remaining tool life include the root mean square (rms) of the minimum power, the mean (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 acquiring tool image features and three BiLSTM models for learning the sensitivity of tool side edge wear width (VB), spindle vibration signal, and power signal to tool service life. An artificial neural network is used to process the fused features from the three BiLSTM models and the CNN outputs to output the predicted tool service life. Each BiLSTM model is constructed using stacked two-layer BiLSTM networks, with the same sequential structure but different hyperparameters. The hyperparameters are obtained using the Hyperband algorithm, which is based on the Successive Halving algorithm, introducing a elimination mechanism into the grid search method to obtain the optimal hyperparameter combination, including:

[0033] Allocate equal resources to n models with different hyperparameters and the same architecture, train them under the same conditions for round A, eliminate a number of models, and then train them for round B, eliminating a number of models, until only 1 model remains, thus obtaining the optimal combination of hyperparameters.

[0034] In the tool remaining service life prediction model, when processing data, the tool wear area image, VB value, vibration and power signal collected in real time are processed by a 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 wear area image, VB value, vibration and power signal. Then, the artificial neural network determines the tool remaining service life based on the fusion features.

[0035] This invention focuses on the remaining service life of cutting tools. Based on an improved CNN-BiLSTM integrated network as the backbone, it uses the image features of the tool wear zone, the wear width VB value of the tool side edge obtained from the image features, and the vibration and power signal features of the spindle during machining as the basis for prediction. It can comprehensively and intuitively monitor the tool wear status, achieving highly real-time and accurate prediction. It has good predictive ability in the early stage of tool wear, with an average prediction error of 4.76 min and a prediction time of less than 1 second. Attached Figure Description

[0036] Figure 1 This is a flowchart illustrating the implementation of the tool health status prediction method based on fused information in an embodiment of the present invention.

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

[0038] Figure 3 This is a flowchart illustrating the data acquisition process for tool wear throughout its entire lifecycle, as described in an embodiment of the present invention.

[0039] Figure 4 This is an original view of the side edge of the cutting tool in an embodiment of the present invention.

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

[0041] Figure 6 This is a diagram showing the straight line position of the cutting edge of the tool in an embodiment of the present invention.

[0042] Figure 7 This is a grayscale image of the wear area on the side edge of the cutting tool in 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] Figure 9 The diagram shows the structure of a BiLSTM model for learning the tool wear (VB), vibration, and power sensitivity characteristics in this embodiment of the invention.

[0045] Figure 10 This is the overall block diagram of the CNN-BiLSTM model based on fused data in this invention.

[0046] Figure 11 The diagram below shows the logic block diagram of the Hyperband algorithm for selecting hyperparameters of the CNN-BiLSTM model based on fused data in this invention.

[0047] Figure 12 This is a diagram showing the prediction performance of the CNN-BiLSTM model based on fused data in this invention. Detailed Implementation

[0048] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0049] With the development of computer technology, chip manufacturing technology, and other technologies, artificial intelligence networks have brought a new possibility to the monitoring of tool wear stages and the prediction of remaining useful life. The general approximation theorem states that a neural network with at least one hidden layer can fit an arbitrary function with arbitrary precision. Furthermore, using features extracted by the neural network itself, compared to traditional manual features, can more accurately and comprehensively reflect the tool wear condition. Therefore, a properly trained neural network has the potential to surpass traditional research, more accurately fitting the tool wear degradation process and making more precise and reliable predictions of tool wear condition and remaining useful life (RUL).

[0050] Therefore, the inventors of this application, through creative work, have proposed a tool health status prediction method based on fused information, specifically a method for multi-source data fusion and tool remaining service life prediction based on an improved CNN-BiLSTM model. This method utilizes multi-source information collected during milling, including images, vibrations, and power, to accurately predict the remaining service life of the tool, thereby guiding tool sharpening and replacement, achieving the goals of reducing manufacturing costs, improving processing efficiency, and ensuring processing accuracy.

[0051] To achieve the present invention, the embodiments of the present invention first designed a data acquisition system for tool remaining service life related characteristic data to collect relevant data. Then, based on the data acquisition of image, vibration, and power information, a tool life cycle wear database was established. Next, the acquired raw data was preprocessed to remove redundant information from the multi-source information, and the wear area image of the end mill side edge was extracted based on the method of line recognition and double threshold segmentation, realizing the basic complete extraction of the side edge wear area. Then, based on the correlation analysis method, the sensitive characteristics of vibration and power signals on the remaining tool service life were determined, and the tool failure was judged based on the determined tool failure criteria.

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

[0053] In this embodiment of the application, the preprocessing of the acquired raw data is a method to improve the model fitting accuracy and training speed. The preprocessing includes the preprocessing of the acquired tool wear area image, the vibration of the tool spindle, and the power signal.

[0054] In this embodiment, a tool remaining service life prediction model is constructed based on an improved CNN-BiLSTM. The model performs deep learning on the feature information of preprocessed data, uses artificial neural networks to fuse data from different sources, and finally predicts the tool remaining service life based on the fused information.

[0055] like Figure 1 As shown in the embodiments of this application, a tool health status prediction method based on fused information includes the following steps:

[0056] S1: First, collect the tool wear images, spindle vibration information, and power information during the machine tool processing, which are necessary for the tool remaining service life prediction model. The total service time of the tool from brand new to failure is also collected. Based on the total service 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, appropriate data preprocessing methods are used to remove useless information and enhance effective information. Through preprocessing of tool wear images, spindle vibration information, and power information, normalized tool wear area images, tool side edge wear width VB value, spindle vibration sensitivity to remaining tool service life, and power sensitivity to remaining tool service life are obtained, totaling four target features.

[0058] S3: The four target features mentioned above, together with the remaining tool life obtained from the previous experiment, are used as input data. The preprocessed tool wear area image, the wear width VB value of the tool side edge, the spindle vibration sensitivity to the remaining tool life, and the power sensitivity to the remaining tool life are used as the prediction features. The remaining tool life is used as the prediction target. The constructed tool life prediction model is trained and verified with the validation set.

[0059] After training, the tool remaining service life prediction model outputs a tool remaining service life prediction value with sufficient accuracy based on the preprocessed four feature data.

[0060] S4: During tool prediction, the test set data is processed to obtain four target features, which are then input into the trained tool remaining service life prediction model to obtain the predicted value of the tool's remaining service life.

[0061] In this embodiment of the application, the obtained tool remaining service life prediction model, i.e., the RUL (remaining tool life) prediction model, is as follows:

[0062] Tool wear generally falls into three stages: initial wear, normal wear, and rapid wear. When a tool enters the rapid wear stage, the contact area between the tool flank and the machined surface is large, and both cutting temperature and cutting force increase rapidly. Therefore, instability during machining increases sharply, while machining accuracy and production efficiency decrease rapidly. In production, to ensure machining quality, tools should be replaced before rapid wear occurs. Tool wear mainly occurs on the side cutting edge, i.e., the flank face during side milling. Therefore, in this application, tool data acquisition and subsequent prediction primarily focus on the side cutting edge wear of the end mill.

[0063] In the experiment, a lean four-axis machining center of Tianjin University was used. A HIOKI PW3360 clamp-on power meter was used to measure the power signal of the machine tool, and the average, minimum, and maximum power values 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 2000 Hz. 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. An annular adjustable light source was used to provide illumination to ensure stable light and reduce the difficulty of subsequent processing. After calibration, the width of each pixel in the captured image is 2.082031 μm.

[0064] Table 1

[0065]

[0066] Among them, there is a controllable magnetic adsorption device at the bottom of the CCD industrial camera. During photography (non-machining stage), it is installed at a predetermined position. After installation, the workbench can be manually moved to a fixed photography position to take pictures of the tool. Since the position of the cutting edge after machining is uncontrollable, the four cutting edges of the end mill are manually marked and distinguished on the tool holder with a marker pen. At the same time, marks are made on the display screen to ensure the approximate position of the tool wear area and reduce the difficulty of later processing. The vibration sensor is adsorbed on the spindle to collect the active vibration signal.

[0067] The experimental data was obtained by repeatedly performing straight-line milling on 45 steel with dimensions of 60 mm × 60 mm × 40 mm using a high-speed steel end mill with a diameter of 10 mm. The side-edge image of the end mill was collected once every 6 tool feeds, and the machining path for 6 tool feeds is as Figure 2 shown. The high-speed steel end mill moves in a snake-shaped path when machining on 45 steel. The VB value was manually measured using the supporting software S-EYE of the CCD industrial camera, and the cutting parameters are shown in Table 2 below.

[0068] Table 2

[0069]

[0070] In the experiment, the acquisition process of the wear information of each tool is Figure 3 shown. First, the acquisition of vibration and power signals starts simultaneously with the start of machining. After completing 6 tool feeds along the path, the acquisition of vibration and power signals is paused, and the current time point is marked. Then, the side-edge images of each cutting edge are taken through the imaging system, and the VB value of each cutting edge is manually measured using the software. If the VB value of a cutting edge is greater than 300 μm, the tool is considered to be失效, and the experiment is stopped. Otherwise, the above steps are repeated.

[0071] The above describes the steps and methods for acquiring image, vibration, and power signals in the embodiments of this application. The above is merely one embodiment; the acquisition equipment or methods used in implementing this invention are not limited to those mentioned, as long as they can acquire the spindle vibration signals, power signals, and tool wear images required by this invention.

[0072] In this embodiment of the application, the original image of the side cutting edge of the end mill, obtained under experimental conditions, is shown below. Figure 4 As shown. Much of the content in the original image of the side cutting edge, including the background, is redundant information. It does not help assess the wear condition of the tool or predict its remaining service life; instead, it may become interfering information affecting the prediction process. Therefore, in the embodiments of this application, after acquiring the tool image, available processing methods or techniques are used to extract the required features from the wear area of ​​the tool to obtain the required wear area image of the side cutting edge, and the wear width VB value of the side cutting edge is calculated and normalized. The implementation steps include:

[0073] S21. Convert the initial or original image of the tool's side edge from the RGB color gamut to the HSV color gamut, perform initial binarization thresholding segmentation in the HSV color gamut, and extract the tool cutting edge contour image.

[0074] The formula for converting 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 gamut, and H, S, and V correspond to the three-channel color values ​​in the HSV color gamut. By setting an appropriate threshold in the HSV color gamut, the cutting edge contour image of the tool can be initially extracted, such as... Figure 5 As shown, Figure 5 This is the preliminary extracted image of the cutting edge contour of the tool.

[0078] S22. In Figure 5 Based on the tool cutting edge contour image shown, straight line recognition is performed to extract the location of the tool cutting edge and the cutting edge slope k0. Then, secondary thresholding 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 in a rectangular coordinate system can be represented as: y = kx + b, where k and b are parameters representing the slope and intercept of the line. All lines passing through a point A(x_0, y_0) satisfy the equation y_0 = kx_0 + b, meaning point A(x_0, y_0) defines a family of lines. Rewriting the equation as: b = -x_0k + y_0, the family of lines defined by point A(x_0, y_0) corresponds to a single line in the parameter space kb. The line defined by points A(x_0, y_0) and B(x_1, y_1) corresponds to a unique point in the parameter space kb; that is, the family of lines defined by point A corresponds to the intersection of line l_A and the family of lines defined by point B in the parameter space kb.

[0080] Therefore, by finding the point (k_0, b_0) where all points X_i on the image correspond to the lines l_i in the parameter space kb with the most intersecting lines, we can determine the line y = k_0x + b_0 where the blade lies. Figure 5 The straight line obtained based on the above is as follows Figure 6 As shown by the red line, shifting this line left and right, retaining only the enclosed area, yields the blade position. Then, a second thresholding segmentation is performed in the HSV color gamut, selecting an appropriate threshold to obtain the image of the side blade wear area. Based on this, the image is converted to grayscale to reduce the data volume, such as... Figure 7 As shown.

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

[0082] The aforementioned steps have yielded the blade slope k0. The tilt angle θ0 and the counterclockwise angle θ of the image are obtained from the following formula:

[0083]

[0084] θ = 90 - θ0

[0085] Based on the relationship between the tilt angle θ0 and the counterclockwise angle θ of the image, the side edge wear area image 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] In the process of rotating the image, the image is first rotated 45 degrees clockwise, and its diagonal is used as the length and height of the new image. The missing data in the new image is filled with 0 values, i.e., black. Then the new image is rotated θ+45 degrees counterclockwise to avoid the loss of local information.

[0087] With the cutting edge in a vertical position, the pixels are calibrated, and the wear width VB value of the side edge is calculated using the following formula:

[0088]

[0089] VB = max(b1, b2, b3, ..., b n )×μ

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

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

[0092] The grayscale image of the side edge wear area can be represented as follows:

[0093] GWI r =[p ij ] 1080×1920

[0094] In the formula, p ik This represents the grayscale value of each point in the side edge wear zone. (GWI) r This represents a grayscale image of the side blade 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 grayscale image normalization process of the side blade wear area can be simply represented as:

[0096]

[0097] In the formula, GWI s This 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 The maximum value among all VB values ​​is obtained based on the data. MIN It is the minimum value among all VB values ​​obtained from the data. is VB represents the normalized wear width of the tool side edge. i This represents the wear width VB value at different positions on the tool side edge before normalization.

[0101] Various characteristics in the time domain can intuitively reflect the instantaneous changes of a signal. Among them, the mean (mv), root mean square (rms), and standard deviation (std) reflect the amplitude and central tendency of the signal, while the waveform factor (sf), skewness (ske), and kurtosis (kur) reflect the waveform of the signal. Tool wear is also reflected in the spindle vibration and the frequency characteristics of the power signal. Therefore, this application selects three typical frequency domain characteristics—average power spectrum (mps), frequency centroid (fc), and mean square frequency (msf)—to reflect its power spectrum, centroid position, and dominant frequency band, respectively.

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

[0103] After denoising the original acquired spindle vibration and power data, the characteristic values ​​of the vibration and power signals are obtained. Then, the correlation analysis of the remaining tool life is performed to eliminate features that are not sensitive to the remaining tool life, extract the sensitive features that are sensitive to the remaining tool life, and normalize them to obtain the sensitivity features of spindle vibration to the remaining tool life and the sensitivity features of machine tool power to the remaining tool life.

[0104] Since not all the desired eigenvalues ​​are affected by tool usage time or remaining tool life, a correlation analysis is performed on the characteristics of the spindle vibration and power signals with the remaining tool life to eliminate features insensitive to remaining tool life. Specifically, this can be achieved through the following steps:

[0105] S31. Calculate the correlation coefficients between each eigenvalue and tool life:

[0106] The formula for calculating the correlation coefficient is expressed as follows:

[0107]

[0108] In the formula, r represents the correlation coefficient between each characteristic value and the tool life, and x and y represent different types of characteristic values.

[0109] S32. Construct a dimensionless index for the correlation coefficients between the obtained eigenvalues ​​and tool life.

[0110] Since each signal involves data from multiple cutting tools, and the data differs between each tool, the correlation coefficient alone cannot quantify the performance of each feature. Therefore, a dimensionless index is constructed to measure the overall correlation of each feature value with the tool life. This dimensionless index serves as the standard for selecting sensitive features. The expression for the dimensionless index is as follows:

[0111]

[0112] range = rmax -r min

[0113]

[0114] Where, coef represents a dimensionless index, mean and std represent the mean and standard deviation of the correlation coefficients for each tool under each characteristic value, respectively, and r max With r min These represent the maximum and minimum values ​​of the correlation coefficient r for each tool under each characteristic value.

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

[0116] Specifically, among the power signals, the features of the minimum power value have the best correlation with the remaining tool life, with the root mean square (RMS), mean value (mV), standard deviation (std), average power spectrum (mpS) in the frequency domain, and waveform factor (sf) being the optimal. Similarly, among the vibration signal features, the mean value (mV), root mean square (RMS), mpS, standard deviation (std), and waveform factor (sf) have the best correlation. Therefore, the five sensitive features of the power signal are constructed using the root mean square (RMS), mean value (mV), standard deviation (std), average power spectrum (mpS) in the frequency domain, and waveform factor (sf). Similarly, the five sensitive features of the vibration signal are constructed using the mean value (mV), root mean square (RMS), average power spectrum (mpS) in the frequency domain, standard deviation (std), and waveform factor (sf). The correlation coefficient (coef) of vibration value, maximum power value, minimum power value, and average power value is shown in Table 3 below.

[0117] Table 3

[0118]

[0119] S34. Normalize each sensitive feature within its scope based on the entire dataset to address the inconsistency in magnitude and dimension between different sensitive features:

[0120]

[0121] Among them, X MAX X is the maximum value among all such sensitive features obtained from the data. MIN It is the minimum value among all such sensitive features obtained from the data, X i X represents the data representing the sensitive features to be processed. is This represents data representing sensitive features after normalization.

[0122] In this embodiment of the 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 Networks (CNNs) are neural networks used for image recognition and prediction. Compared to classic feedforward neural networks, CNNs significantly reduce the number of network parameters, improve learning and inference speed, and possess translation invariance. Furthermore, their rotation and scaling invariance can be enhanced using image augmentation techniques. A typical CNN network structure consists of convolutional layers, pooling layers, and fully connected layers. Figure 8 As shown.

[0124] The forward propagation algorithm for convolutional layers is shown below:

[0125]

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

[0127] In the formula, X (l-1) For the input tensor, W (l) For the convolution kernel tensor, b (l) For the bias tensor, Z (l) Let X be the net output tensor, a(·) be the activation function, and X be the net output tensor. (l) To output a tensor, the symbol This 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 is learned through backpropagation. The CNN network structure used in this invention to extract tool image features 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 consists of the following layers in sequence: Input layer, Conv layer (using ReLU activation function), Maxpooling layer, Conv layer (using ReLU activation function), Averagepooling layer, Flatten layer, Dropout layer, and Dense fully connected layer (using Leaky ReLU activation function). The Dense fully connected layer is implemented using an artificial neural network, which combines the tool image features x from different positions output by the previous layer. nAfter dimensionality reduction processing, the dimensionality-reduced feature h is output. n Finally, the target features of the tool wear area image obtained after learning are obtained.

[0131] In this application, the tool remaining service life prediction model learns the sensitivity features of VB value, spindle vibration signal, and power signal based on a BiLSTM model. Specifically, for the sensitivity features of VB value, spindle vibration signal, and power signal, this invention uses a Bidirectional Long Short-Term Memory (BiLSTM) neural network. LSTM networks are a type of Recurrent Neural Network (RNN). Compared to S-RNNs, LSTM networks introduce memory elements and gating mechanisms, exhibiting stronger processing capabilities for long-distance dependencies and effectively avoiding the gradient vanishing and gradient exploding problems present in S-RNNs. BiLSTM networks are constructed using bidirectional LSTM networks, allowing for a more comprehensive consideration of contextual information. Its basic network structure is as follows: Figure 9 As shown, the three input values ​​X1, X2, and X3 are input into the network and processed by two layers of bidirectional BiLSTM network before being output.

[0132] This invention employs a stack of two bidirectional BiLSTM networks combined with a reshape layer to extract deeper information. The network structure for processing VB values, vibration signals, and power signal sensitivity features is shown in Table 5 below.

[0133] Table 5

[0134]

[0135] As shown in the table above, the two-layer bidirectional BiLSTM network sequentially includes an input layer, a reshape layer, a BiLSTM layer, another reshape layer, and another BiLSTM layer. It processes the sensitive features of the VB value, vibration signal, and power signal. The input VB value is a single feature sequence, while the vibration signal and power signal each have five feature sequences. Through layer-by-layer deep learning using the bidirectional long short-time neural network, the final output includes 10 feature sequences for the VB value, 34 feature sequences for the vibration signal, and 30 feature sequences for the power signal.

[0136] Therefore, in this embodiment of the application, the overall network structure of the tool remaining service life prediction model is as follows: Figure 10As shown, this is a CNN-BiLSTM model, consisting of a CNN network and a BiLSTM network. For the tool wear area image, the CNN network performs deep feature learning on the side edge wear area image. For VB (Vibration Value), vibration sensitivity to remaining service life, and power sensitivity to 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 model's forward inference, the sensitivity features of the wear area image, VB value, vibration, and power signal to the tool's remaining service life are passed through various parts of the backbone network to the head network, where they are nonlinearly fused into several abstract fused features X. n Each abstract fusion feature X n Each feature is composed of wear zone images, VB values, vibration and power signals, and dimensionality reduction features H are formed through a Dense layer. n Finally, the predicted lifetime Y is output.

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

[0138] Feature fusion is achieved through the backpropagation process of the prediction model, where the model learns on its own: the error between the remaining tool life predicted by forward inference and the actual value will be transmitted to various parts of the model during the backpropagation process, thereby allowing the model to automatically adjust its parameters and ultimately determine the composition of the fused features.

[0139] The hyperparameters of the tool remaining service life prediction model include the kernel shape and activation function of each convolutional layer; the kernel size of each convergence layer; the masking ratio of the dropout layer; the number of units and activation function of each fully connected (Dense) layer; the number of units and sequence length of each BiLSTM layer; the number of training epochs; the batch size; and the model learning rate. The Hyperband algorithm is used in the hyperparameter optimization process. This algorithm, based on the Successive Halving algorithm, introduces a elimination mechanism into the grid search method to obtain the optimal hyperparameter combination: that is, allocating equal resources to n different hyperparameter models with the same architecture, training them under the same conditions for epoch A, eliminating a certain number of models, and then further training them for epoch B, eliminating a certain number of models, until only 1 model remains, thus obtaining the optimal hyperparameter combination. The program flow is as follows: Figure 11 As shown.

[0140] In this embodiment, 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 verify the model's performance after each training round and does not participate in the training itself; the test set is used to verify the model's generalization prediction ability after training is complete. The full-cycle prediction results for the collected six tools are as follows: Figure 12 As shown.

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

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

[0143] Table 6

[0144]

[0145] Among the image, vibration, and power indicators, the prediction performance increases in that order. Compared to using vibration or power alone, fusing vibration and power signals reduces MAE by 6.92 and 0.3, MSE by 377.79 and 18.6, and the maximum absolute error by 15.94 and 2.2, respectively. Adding image data further reduces MAE by 3.21 and MSE by 83.31 compared to using only vibration and power signals. Fusing image, VB, vibration, and power information yields the optimal MAE, MSE, and maximum absolute error: 4.76, 35.99, and 19.62, respectively. This performance surpasses any of the aforementioned combinations or the use of only VB.

[0146] Wherein, MAE and MSE are the mean absolute error and mean squared error, respectively, and their calculation formulas are as follows:

[0147]

[0148] In the formula, y i For the actual RUL of the i-th data set, Let RUL be the model's prediction for the i-th data set, and n be the total number of data sets in the test set.

[0149] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. It will be apparent to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and that 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 regarded as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, it is intended that all variations falling within the meaning and scope of the equivalents of the claims be included within the invention.

[0151] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A tool health status prediction method based on fused information, characterized in that, Including the following steps: The machine tool wear images, spindle vibration information, and machine tool power information during the machining process are collected to train the tool remaining service life prediction model. The total service time of the tool from brand new to failure is also collected. The remaining service life of the tool at any time is obtained from the total service time of the tool. Preprocessing of tool wear images, spindle vibration information, and power information yields normalized tool wear area images, tool side edge wear width VB value, spindle vibration sensitivity to remaining tool life, and machine tool power sensitivity to remaining tool life, for a total of four target features. Using the target features as the features used for prediction and the remaining tool life as the prediction target, a tool life prediction model is trained using training set data and then verified using a validation set. After processing the test set data, the four target features are input into the trained tool remaining service life prediction model to obtain the predicted value of the tool's remaining service life. The tool remaining service life prediction model includes a convolutional neural network (CNN) for acquiring tool image features and three BiLSTM models for learning the sensitivity features of tool side edge wear width (VB), spindle vibration signal, and machine tool power signal to tool service life. An artificial neural network is used to process the fused features from the three BiLSTM models and the CNN output to output the predicted tool service life value. Each BiLSTM model is constructed using stacked two-layer BiLSTM networks, with the same sequential structure but different hyperparameters. The hyperparameters are obtained using the Hyperband algorithm, which is based on the Successive Halving algorithm, introducing a elimination mechanism into the grid search method to obtain the optimal hyperparameter combination, including: Allocate equal resources to n models with different hyperparameters and the same architecture, train them for round A under the same conditions, eliminate a number of models, and then train them for round B, eliminating a number of models, until only 1 model remains, thus obtaining the optimal combination of hyperparameters. When processing data, the tool remaining service life prediction model uses real-time collected tool wear area images, VB values, vibration and power signals as sensitive features to tool remaining service life. These features are processed by a convolutional neural network (CNN) and three 2*BiLSTM networks, and then nonlinearly fused into several abstract fusion features. Each abstract fusion feature is composed of wear area images, VB values, vibration and power signals. Then, an artificial neural network determines the tool remaining service life based on the fusion features.

2. The tool health status prediction method based on fused information according to claim 1, characterized in that, The wear width VB value of the tool side edge is obtained through the following steps: The original image of the side edge of the tool is converted from the RGB color gamut to the HSV color gamut. Initial binarization thresholding is performed in the HSV color gamut to extract the contour image of the tool cutting edge. Line recognition is performed on the contour image of the cutting edge of the tool to extract the location of the cutting edge and the slope of the cutting edge. Then, secondary thresholding is performed to obtain the side edge wear area image and convert it into a grayscale image. The side edge wear area image is rotated in grayscale form to make the cutting edge vertical, and then the pixels are calibrated to calculate the wear width VB value of the side edge. The wear width VB value data of the side edge is normalized to obtain the normalized wear width VB value of the tool side edge.

3. The tool health status prediction method based on fused information according to claim 2, characterized in that, When rotating the grayscale image of the side edge wear area, firstly, based on the obtained blade slope, determine the relationship between the image's tilt angle and the angle to be rotated counterclockwise. Then, based on this relationship, rotate the image so that the cutting edge in the image is in a vertical state. During rotation, first rotate the image clockwise by 45 degrees, then use its diagonal as the length and height of the new image, filling the missing data in the new image with 0 values ​​(i.e., black). Subsequently, rotate the new image counterclockwise. Spend, The new image needs to be rotated counterclockwise.

4. The tool health status prediction method based on fused information according to claim 3, characterized in that, The wear width VB value of the side blade is calculated using the following formula: ; ; In the formula, , The new image after rotation is the first one. All points in the row The maximum and minimum values ​​of the coordinates. For the wear zone The pixel width of the row, For the new image height, This is the new image scale, which is the actual width of 1 pixel.

5. The tool health status prediction method based on fused information according to claim 4, characterized in that, The wear width VB value of the tool side edge is normalized based on all data, and the expression is as follows: ; in, It is the maximum value in all VB. It is the minimum value in all VB. This represents the normalized wear width VB value of the tool side edge. This represents the wear width VB value of the tool side edge before normalization.

6. The tool health status prediction method based on fused information according to claim 1, characterized in that, The acquisition of the sensitivity characteristics of spindle vibration to remaining tool life and the sensitivity characteristics of machine tool power to remaining tool life includes: After denoising the original acquired spindle vibration and power data, the characteristic values ​​of the vibration and power signals are obtained. Then, the correlation between the characteristic values ​​and the remaining tool life is analyzed to obtain the correlation coefficient. Based on the correlation coefficient, a dimensionless index is constructed. Based on the dimensionless index, features that are not sensitive to the remaining tool life are excluded, and sensitive features that are sensitive to the remaining tool life are extracted. Then, the data are normalized to obtain the sensitive features of spindle vibration on the remaining tool life and the sensitive features of machine tool power on the remaining tool life.

7. The tool health status prediction method based on fused information according to claim 6, characterized in that, The expression for constructing the dimensionless index based on the correlation coefficient is as follows: ; ; ; in, Indicates a dimensionless index. and These represent the mean and standard deviation of the correlation coefficients for each tool under each eigenvalue. and These represent the correlation coefficients of each tool under each eigenvalue. The maximum and minimum values.

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

Images