Real-time tool wear monitoring and predicting method and system based on multi-sensor information fusion

Abstract

The invention belongs to the related technical field of numerical control machining, and discloses a tool wear real-time monitoring and prediction method and system based on multi-sensor information fusion, and the method comprises the steps: S1, collecting a monitoring signal and a tool wear image in a machining experiment process; s2, carrying out wear region segmentation on the tool wear image by adopting a wear region segmentation model, and calculating to obtain a real wear value; s3, establishing a monitoring training sample set by using the monitoring signal and the real wear value, and training the wear monitoring model; establishing a prediction training sample set by using the sequence of the real wear values, and training a wear prediction model; s4, the real-time monitoring signal and the wear monitoring model are used for monitoring the wear value of the tool in real time; and performing online prediction on the wear value of the cutter by using the wear prediction model. According to the invention, the wear value of the cutter can be monitored in real time according to the real-time monitoring signal, online wear value prediction can be realized, and the production efficiency can be greatly improved.

Description

Technical Field

[0001] The present invention belongs to the technical field related to numerical control machining, and more specifically, relates to a real-time monitoring and prediction method and system for tool wear based on multi-sensor information fusion. Background Art

[0002] In mechanical processing, 70%-80% of parts are formed by CNC machine tools, and the tool, as the execution end of the CNC machine tool, will inevitably produce friction with the chips and workpieces. Affected by the cutting force and cutting heat, the tool will continue to wear, and the wear state will gradually deteriorate over time and eventually break the edge, seriously affecting the processing efficiency, processing quality and processing cost. Previous studies have shown that 10%-40% of the total machine downtime is caused by tool failure. In addition, data show that about 75% of cutting failures are due to the use of severely worn cutting tools. Therefore, intelligent monitoring and prediction of tool wear is of great significance for establishing a reasonable tool change strategy and ensuring the continuity and safety of the processing process.

[0003] At present, the tool wear monitoring and prediction methods are divided into two categories: direct method and indirect method. The direct method directly measures the geometric dimensions of tool wear through microscopes, CCD cameras, etc., and its monitoring method is more intuitive and has higher accuracy. However, this method requires shutdown monitoring, and it is difficult to monitor tool wear in real time due to the influence of processing environment such as chips and cutting fluid. In contrast, the indirect method indirectly reflects the tool wear state by analyzing signals such as vibration, cutting force, acoustic emission, and current that are strongly related to tool wear. This indirect wear measurement method is more robust and can be monitored in real time, so it is used more frequently. Due to the complexity and nonlinearity of tool wear in actual cutting processes, as well as the diversity of materials and processes, it is difficult to establish an accurate tool wear mechanism model. At this stage, the indirect method usually uses a data-driven intelligent method, that is, the relationship between sensor signal characteristics and tool status is established through machine learning or even deep learning methods.

[0004] However, there are the following problems in existing research: when constructing data sets in most studies, the label data of tool wear values ​​are obtained through offline measurement. This method is time-consuming and labor-intensive, and it is impossible to obtain a large number of samples, which hinders the research progress. Summary of the invention

[0005] In response to the above defects or improvement needs of the prior art, the present invention provides a real-time monitoring and prediction method and system for tool wear based on multi-sensor information fusion, which is used to solve the problem that in the existing research on monitoring the tool status through machine learning methods, when constructing a data set, the label data of the tool wear value needs to be obtained through measurement under offline conditions, which is time-consuming, labor-intensive and inefficient.

[0006] To achieve the above object, according to one aspect of the present invention, a method for real-time monitoring and prediction of tool wear based on multi-sensor information fusion is provided, comprising:

[0007] Offline stage:

[0008] S1, conduct a machining experiment, collect monitoring signals obtained by multiple sensors and tool wear images obtained by a camera during the machining experiment;

[0009] S2, using a trained wear area segmentation model to segment the wear area of ​​the tool wear image, and calculating and obtaining a true wear value based on the segmented wear area;

[0010] S3, using the monitoring signal and the corresponding real wear value to establish a monitoring training sample set, and using the monitoring training sample set to train a wear monitoring model to obtain a trained wear monitoring model;

[0011] Using the sequence of real wear values ​​to establish a prediction training sample set, using the prediction training sample set to train a wear prediction model, and obtaining a trained wear prediction model;

[0012] Online stage:

[0013] S4, during the machining process, using multiple sensors to obtain real-time monitoring signals; using the real-time monitoring signals and the trained wear monitoring model to monitor the wear value of the tool in real time;

[0014] The trained wear prediction model is used to predict the wear value of the tool.

[0015] According to the real-time monitoring and prediction method for tool wear based on multi-sensor information fusion provided by the present invention, before training the wear monitoring model using the monitoring training sample set in S3, the method further includes:

[0016] Performing noise reduction and feature extraction processing on the monitoring signal to obtain preprocessing features; when training the wear monitoring model, the preprocessing features are used as inputs of the wear monitoring model, and the output of the wear monitoring model is a predicted value of the wear value;

[0017] The noise reduction process on the monitoring signal includes:

[0018] Performing at least one of average filtering, wavelet transform, inverse wavelet transform and dead-load removal processing on the monitoring signal;

[0019] Performing feature extraction processing on the monitoring signal includes:

[0020] Multi-domain features are extracted from the time domain, frequency domain, and time-frequency domain of the monitoring signal after noise reduction processing; the wear strong correlation features are screened out from the multi-domain features using the Pearson correlation coefficient method; the high-dimensional features of the wear strong correlation features are mapped to the low-dimensional space using the KPCA method to obtain preprocessing features.

[0021] According to the real-time monitoring and prediction method for tool wear based on multi-sensor information fusion provided by the present invention, in the feature extraction processing of the monitoring signal, before mapping the high-dimensional features of the wear-strong correlation features to the low-dimensional space using the KPCA method, it also includes:

[0022] An exponential filtering operation is performed on the wear-strongly correlated features.

[0023] According to the real-time monitoring and prediction method for tool wear based on multi-sensor information fusion provided by the present invention, the monitoring signal is subjected to noise reduction processing and feature extraction processing to obtain preprocessing features, and further includes:

[0024] The monitoring signals acquired by multiple sensors are subjected to noise reduction processing and multi-domain feature extraction respectively, and then the multi-domain features corresponding to the monitoring signals acquired by multiple sensors are fused, and then the Pearson correlation coefficient method is used to screen out the wear-related features based on the fused multi-domain features.

[0025] According to the real-time monitoring and prediction method for tool wear based on multi-sensor information fusion provided by the present invention, the wear area segmentation model in S2 is a YOLOv8 segmentation model; the wear area of ​​the tool wear image obtained in S1 is labeled, a segmentation sample set is constructed, and the wear area segmentation model is trained using the segmentation sample set.

[0026] According to the real-time monitoring and prediction method for tool wear based on multi-sensor information fusion provided by the present invention, in S2, the actual wear value is calculated based on the segmented wear area, specifically including:

[0027] Obtain pixel coordinates of vertices of the wear area to form a vertex point set;

[0028] Traversing the vertex point set, obtaining equations of line segments between two vertices; and determining the original cutting edge line segment of the tool by screening the slope of the line segment and the number of vertices passed;

[0029] Calculate the vertical distances between the remaining points of the vertex point set and the original cutting edge line segment of the tool, and select the value with the largest vertical distance as the number of pixels of the maximum wear width;

[0030] According to the number of pixels of the maximum wear width and the size of the tool wear image, the actual value of the maximum wear width is obtained as the true wear value.

[0031] According to the real-time monitoring and prediction method of tool wear based on multi-sensor information fusion provided by the present invention, the multiple sensors in S1 include multiple vibration sensors, cutting force sensors and acoustic emission sensors; the wear monitoring model in S3 is a KAN deep learning model; and the wear prediction model in S3 is a Transformer model.

[0032] According to the real-time monitoring and prediction method for tool wear based on multi-sensor information fusion provided by the present invention, the wear value of the tool is predicted using the trained wear prediction model, which specifically includes:

[0033] The wear prediction model inputs a monitoring value sequence of tool wear values ​​when predicting for the first time, and outputs a prediction value sequence of tool wear values;

[0034] Afterwards, the wear prediction model inputs a combination of the monitoring value of the tool wear value and the predicted value output from the previous prediction, or a sequence of predicted values ​​output from the previous prediction, and outputs a sequence of predicted values ​​of the tool wear value, forming an advanced prediction mode.

[0035] The monitoring value of the tool wear value is the output value of the wear monitoring model.

[0036] According to the real-time monitoring and prediction method for tool wear based on multi-sensor information fusion provided by the present invention, the advance prediction mode also includes:

[0037] Timely correction strategy: comparing the predicted value sequence of tool wear values ​​output by the wear prediction model with the corresponding monitored value sequence of tool wear values ​​to obtain the error between the predicted value and the monitored value;

[0038] If the error is less than the preset threshold, online prediction will continue in the advance prediction mode; if the error is greater than or equal to the preset threshold, when the wear prediction model makes the next prediction, the monitoring value sequence of the tool wear value will be input to correct the model input; thereafter, online prediction will continue in the advance prediction mode.

[0039] According to another aspect of the present invention, a real-time monitoring and prediction system for tool wear based on multi-sensor information fusion is provided, comprising a memory and a processor, wherein the memory stores a computer program, and the processor executes any one of the above-mentioned methods for real-time monitoring and prediction of tool wear based on multi-sensor information fusion when executing the computer program.

[0040] In general, compared with the prior art, the above technical solutions conceived by the present invention provide a method and system for real-time monitoring and prediction of tool wear based on multi-sensor information fusion:

[0041] 1. The trained wear area segmentation model can be used to achieve online segmentation of the tool wear area, and then the wear value can be calculated online. There is no need to obtain the wear value through offline testing, which can significantly improve the acquisition speed of the monitoring model and prediction model training sample set and simplify the operation; further training and establishing a wear monitoring model can realize real-time monitoring of the tool wear value according to the real-time monitoring signal, and then the wear prediction model obtained based on the wear value training can be used for online prediction, which can realize online wear value prediction and real-time prediction without interfering with production. This prediction method can greatly improve production efficiency;

[0042] 2. Through multi-domain feature extraction and KPCA dimensionality reduction, the most representative and explanatory features are obtained, which not only improves the prediction accuracy but also effectively removes redundant information;

[0043] 3. The tool wear segmentation model based on YOLOv8 can efficiently identify the tool wear area and accurately calculate the tool wear value with an error of less than 5μm;

[0044] 4. The prediction accuracy of the KAN tool wear real-time monitoring model on three sets of experimental data in public data reached 100%, 99.99%, and 99.98% respectively, and reached 99.95% on experimental data. At the same time, the error is extremely low in both public and experimental data;

[0045] 5. Under the traditional prediction mode, the Transformer tool wear prediction model achieved prediction accuracies of 99.98%, 99.94%, and 99.94% on three sets of experimental data from public data, and 99.94% on experimental data; at the same time, it had extremely low errors on both public and experimental data;

[0046] 6. In the advanced prediction mode, the tool wear prediction model generally has poor prediction results under different T1 and T2 parameters. However, by adopting a timely correction strategy, under different T1 and T2 parameters, the three sets of experimental data in the public data have at least 99.31%, 97.00% and 99.14% respectively; the experimental data has at least 99.38% accuracy; at the same time, both the public and experimental data have low errors. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is the overall structure diagram of the real-time monitoring and prediction method of tool wear based on multi-sensor information fusion provided by the present invention;

[0048] Figure 2 This is the structure diagram of the YOLOv8 tool wear area segmentation model;

[0049] Figure 3 is the change of the training indicators of the segmentation model;

[0050] Figure 4 This is the inference effect diagram of the tool wear segmentation model, where (a) is not worn; (b) is the initial stage of wear; (c) is the middle stage of wear; (d) is the late stage of wear;

[0051] Figure 5 It is a process of calculating the quantified tool wear value, wherein: (a) the original cutting edge is identified; (b) the quantified tool wear value is obtained; (c) the conversion between pixel value and actual length is performed;

[0052] Figure 6 It is a spectrum diagram of public data, where (a) force signal spectrum of C1 dataset; (b) force signal spectrum of C4 dataset; (c) force signal spectrum of C6 dataset; (d) vibration signal spectrum of C1 dataset; (e) vibration signal spectrum of C4 dataset; (f) vibration signal spectrum of C6 dataset;

[0053] Figure 7 It is the feature map before and after dimensionality reduction before feature post-processing;

[0054] Figure 8 It is the feature map before and after dimensionality reduction after feature post-processing;

[0055] Fig. 9 This is the principle diagram of the advance prediction mode. DETAILED DESCRIPTION

[0056] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0057] See also Figure 1 This embodiment provides a real-time monitoring and prediction method for tool wear based on multi-sensor information fusion, the method comprising:

[0058] Offline stage:

[0059] S1, conduct a machining experiment, collect monitoring signals obtained by multiple sensors and tool wear images obtained by a camera during the machining experiment;

[0060] S2, using a trained wear area segmentation model to segment the wear area of ​​the tool wear image, and calculating and obtaining a true wear value based on the segmented wear area;

[0061] S3, using the monitoring signal and the corresponding real wear value to establish a monitoring training sample set, and using the monitoring training sample set to train a wear monitoring model to obtain a trained wear monitoring model;

[0062] Using the sequence of real wear values ​​to establish a prediction training sample set, using the prediction training sample set to train a wear prediction model, and obtaining a trained wear prediction model;

[0063] Online stage:

[0064] S4, during the machining process, using multiple sensors to obtain real-time monitoring signals; using the real-time monitoring signals and the trained wear monitoring model to monitor the wear value of the tool in real time;

[0065] The trained wear prediction model is used to predict the wear value of the tool.

[0066] This embodiment further considers that there are few existing studies on tool wear monitoring and prediction. The combination of the two can combine short-term real-time monitoring with long-term planning and decision-making, significantly improving production efficiency, ensuring product quality, and extending tool life.

[0067] In addition, the existing tool wear prediction model belongs to the interpolation type prediction. The model is based on limited historical wear data and makes predictions for a limited time step in the future. This prediction model does not have the ability to perceive the wear trend of a very long time step, and has certain limitations in practical applications. In view of the shortcomings of the above-mentioned prior art, the main solutions of the present invention are:

[0068] Establish a fully automatic data acquisition platform to automatically collect multi-sensor information and tool wear images; extract multi-domain feature information of multi-sensor data, and obtain the most representative and explanatory monitoring model input features through fusion and post-processing, and through effective dimensionality reduction methods; use tool wear images to build a machine vision model to automatically identify tool wear areas, and further calculate and obtain quantified tool wear values ​​to provide label data for the data set; build a tool wear monitoring model to achieve real-time mapping of sensor multi-domain feature data to tool wear values; build a tool wear prediction model to achieve the mapping relationship between historical and future tool wear sequences; improve the traditional prediction model and combine it with timely correction strategies to achieve ultra-long time step prediction while ensuring prediction accuracy.

[0069] Specifically, an automated data acquisition platform is built to automatically acquire multi-sensor data and tool wear images to provide support for feature and label data. When acquiring tool wear images, a safe photo-taking point is set inside the machine tool, and two industrial cameras and lenses are installed below and on the side to acquire the bottom and side images of the tool respectively. In order to obtain the image of each side edge of the tool, the M100CN instruction is developed to orient the machine tool spindle to any angle, where N is the angle of rotation with the encoder zero point of the machine tool as the reference point. In addition, the software supporting the camera provides the function of taking photos in a shortcut key mode. Combined with the FOCAS communication protocol, it can realize the automatic acquisition of the wear images of each edge on the bottom and side of the tool. When acquiring multi-sensor information, by installing vibration sensors, dynamometers, acoustic emission sensors, etc. on the machine tool workbench, connecting the data acquisition card to the PC, and using the shortcut key recording function provided by the Devsoft software, it is also possible to realize the automatic acquisition of multi-sensor data.

[0070] After obtaining the tool wear image, use annotation software to annotate the wear area to generate a label file to construct a data set, and divide it into a training set and a test set according to a certain ratio. Build a deep learning instance segmentation model, use the wear image and its label file as the model input to iteratively train the model, so that it can autonomously learn key features to determine the optimal parameters, and finally have the ability to segment the tool wear area, classify the input image at the pixel level, and intelligently segment the tool wear area. After the training is completed, the performance of the model is evaluated on the test set. In order to obtain the quantitative tool wear value, the quantitative wear value is further calculated based on the morphology on the basis of the instance segmentation of the tool wear area. This value provides the label value for the data set of the subsequent tool wear real-time monitoring and prediction tasks.

[0071] After obtaining multi-sensor information, it is first pre-processed by filtering, wavelet transform, and de-idling. Secondly, multi-domain features are extracted from the time domain, frequency domain, and time-frequency domain, and the most representative features are obtained through feature post-processing, feature fusion and dimensionality reduction, and the Pearson correlation coefficient method is used to screen features that are strongly related to wear. Finally, the KPCA method is used to map high-dimensional features to low-dimensional space, so that the model can focus on truly valuable features. This feature provides characteristic values ​​for the data set of subsequent tool wear real-time monitoring and prediction tasks.

[0072] A KAN (Kolmogorov-Arnold Networks) real-time monitoring model for tool wear is established. The neural network is used to establish the nonlinear relationship between multi-sensor multi-domain features and quantified tool wear values, and the real-time mapping of multi-sensor multi-domain features to tool wear values ​​is completed. First, the data set is divided into a training set and a test set to evaluate the performance of the model during the training process. Secondly, the input features are passed to each neuron in the hidden layer. Each hidden layer neuron calculates the univariate function output of each input variable and sums it, and then performs a nonlinear transformation. Thirdly, the outputs of all hidden layer neurons are weighted and summed to obtain the final output. This process realizes the nonlinear combination of high-dimensional inputs through univariate functions to approximate the output of the target function. Then, in the training stage, all parameters are randomly initialized, and the difference between the model output and the true label is calculated to obtain the loss function. Finally, all parameters of the model are updated with the help of the back-propagation algorithm to obtain the monitoring model with the best performance. Its output sequence is used as the input of the Transformer prediction model to calculate the future tool wear sequence.

[0073] A Transformer tool wear prediction model is established, and a neural network is used to establish a nonlinear relationship between the historical tool wear sequence and the future tool wear sequence to complete the prediction of the tool wear value from history to the future. First, the input feature data and label data are properly processed and formatted, and divided into a training set and a test set. Secondly, the input is reshaped into a three-dimensional format [batch size, sequence length, features]. This is because the Transformer model is designed based on sequence processing, and the time step and feature dimension of the sequence need to be clarified. Then, the model is constructed with the input layer, linear embedding layer, multi-head attention layer, feedforward neural network layer, global average pooling layer, dropout layer and output layer, and the model output is calculated and the loss is calculated with the label. Finally, the parameters of the model are iteratively updated through the back-propagation algorithm to obtain the optimal parameter prediction model.

[0074] In view of the fact that the traditional interpolation prediction mode lacks the ability to perceive wear trends in ultra-long time steps, an advance prediction mode is proposed. In this mode, the model uses the prediction output of the previous time step as input each time it predicts, so it has the prediction ability of infinite time steps. At the same time, the decrease in accuracy caused by the accumulation of prediction errors is becoming more and more serious. Therefore, the present invention proposes a timely correction strategy. The timely correction strategy can make up for the error accumulation problem in the advance prediction mode and correct the model input in time; the combination of the advance prediction mode and the timely correction strategy not only enables the model to have ultra-long time step prediction capabilities, but also ensures the model prediction accuracy.

[0075] In another embodiment, a real-time monitoring and prediction system for tool wear based on multi-sensor information fusion is also provided. The system includes a memory and a processor. The memory stores a computer program. When the processor executes the computer program, it executes the real-time monitoring and prediction method for tool wear based on multi-sensor information fusion described in any embodiment.

[0076] In some specific embodiments, the present invention provides a method for real-time monitoring and advance prediction of tool wear based on multi-sensor information fusion, including a fully automated data acquisition platform, a tool wear segmentation model, a tool wear quantification method, multi-sensor information multi-domain feature extraction, a tool wear real-time monitoring model, a tool wear prediction model, a tool wear advance prediction mode, and a timely correction strategy, such as Figure 1 shown.

[0077] Example 1

[0078] Build the fully automated data acquisition platform; set up a safe photo-taking point inside the machine tool. The photo-taking point can be set outside the processing area of ​​the machine tool, and two industrial cameras are installed below and to the side of the photo-taking point to obtain the bottom and side wear images of the tool respectively;

[0079] On the one hand, in order to automatically obtain the image of each cutting edge of the tool, the M100CN instruction was developed to orient the machine tool spindle to any angle, where N is the angle at which the machine tool spindle rotates. The TCP / IP network protocol was used to establish communication between the machine tool and the PC, and key information such as variables, status, and parameters of the FANUC CNC system was obtained based on the FOCAS communication protocol. In addition, the software supporting the camera has the function of taking pictures and storing images using shortcut keys, which makes it possible to automatically obtain images using C# scripts.

[0080] On the other hand, by installing vibration sensors and dynamometers on the machine tool workbench, combined with data acquisition cards and PCs, multi-sensor information can be collected; the Devsoft software supporting the data acquisition card has the function of storing data with shortcut keys, which also facilitates the automatic acquisition of multi-sensor data using C# scripts;

[0081] The monitoring signal collected in S1 has a corresponding relationship with the tool wear image. At the end of each monitoring signal acquisition, the tool can be controlled to move to the image acquisition point, the bottom wear image of the tool can be acquired through the industrial camera below, the side edge wear image of the tool can be acquired through the industrial camera on the side, and the tool can be controlled to rotate at the image acquisition point to acquire the wear images of different side edges of the tool; wherein, the bottom wear image of the tool and the side edge wear image of the tool form a tool wear image. The tool wear image corresponding to each monitoring signal can be obtained, and then the tool wear value corresponding to each monitoring signal can be obtained. For example, the monitoring signal can be collected during each tool processing, and then the tool wear image can be collected when the tool processing is completed. Then, the wear area of ​​the tool wear image acquired in S1 is annotated, and a segmentation sample set is constructed, and the wear area segmentation model is trained using the segmentation sample set.

[0082] In general, during the machining process, first, the C# script is used to automatically and in real time acquire multi-sensor data; second, the machine tool NC code subroutine is developed to make the tool automatically move to the photo-taking point after a certain number of cuts and then pause the machine tool; then, the wear images of the bottom and side edges of the tool are automatically acquired through the C# script and the NC code; finally, the machining process is continued, and this process is repeated. The fully automated data acquisition is realized with the help of FOCAS communication and C# script. The specific steps are shown in Table 1.

[0083] Table 1 Online measurement method of tool wear

[0084]

[0085]

[0086] Example 2

[0087] The tool wear area segmentation model is constructed as a YOLOv8 segmentation model; the tool wear segmentation model is constructed using the YOLOv8 open source project; in order to identify the tool wear area, the wear area of ​​the tool is segmented using machine vision; the wear area in the image is manually marked using marking software, and the 2176 samples obtained are divided into a training set, a validation set, and a test set in a ratio of 1546:546:84; the training set is used to train the YOLOv8 instance segmentation model, and the validation set and the test set are used to evaluate the instance segmentation performance of the YOLOv8 model;

[0088] Use as Figure 2The YOLOv8 model architecture shown in the figure performs instance segmentation on the wear area of ​​the tool; the model is mainly composed of basic modules such as Conv, C2f, SPPF, Upsample, and Concat; the Conv module performs convolution operations on the feature map to extract local features of the image; the C2f module achieves efficient calculation by splitting, processing, and fusing features; the SPPF module realizes multi-scale extraction and fusion of features through multi-level spatial pyramid pooling; the Upsample module uses deconvolution to restore the resolution of the feature map, thereby improving the positioning accuracy of the target; the Concat module is used for feature fusion and jump connection; these basic modules form the YOLOv8 network architecture according to the connection method on the right side of the figure, and different Segment layers identify large, medium, and small scale targets in the input image; this output layer has anchor frame design of different sizes, which enables the model to realize multi-scale feature extraction and fusion, ensuring good detection of multi-scale targets;

[0089] After configuring the YOLOv8 operating environment on the PC and preparing the tool wear images and corresponding label files, iterative training began. After 200 rounds of iterative training, the indicators of the YOLOv8 model changed as follows: Figure 3 As shown in the figure, it can be seen that the loss in the whole training process gradually decreases to a smaller value. In addition, the precision and recall of BoundingBox and Mask both reach 100%. The mAP50, mAP50:95 of BoundingBox and the mAP50, mAP50:95 of Mask reach 0.995, 0.995, 0.995 and 0.913 respectively. Obviously, the model has a very superior instance segmentation performance.

[0090] like Figure 4 The results of model instance segmentation of randomly selected pictures from various stages of the entire tool life cycle process. It can be seen that the model can effectively identify the wear area at any stage of tool wear, providing a basis for subsequent tool wear measurement.

[0091] Example 3

[0092] In S2, the actual wear value is calculated based on the segmented wear area, specifically including:

[0093] Obtain pixel coordinates of vertices of the wear area to form a vertex point set;

[0094] Traversing the vertex point set, obtaining equations of line segments between two vertices; and determining the original cutting edge line segment of the tool by screening the slope of the line segment and the number of vertices passed;

[0095] Calculate the vertical distances between the remaining points of the vertex point set and the original cutting edge line segment of the tool, and select the value with the largest vertical distance as the number of pixels of the maximum wear width;

[0096] According to the number of pixels of the maximum wear width and the size of the tool wear image, the actual value of the maximum wear width is obtained as the true wear value.

[0097] That is, a tool wear quantification method is proposed, and the tool wear value is further calculated by segmenting the wear area; the International Organization for Standardization stipulates that the blunting standard of cemented carbide tools during rough machining is: under normal wear of cutting tools, the average wear width of the tool back face VB average =0.3 or maximum wear width VB max =0.5 is the blunting standard; the present invention adopts the maximum wear width VB max Conduct follow-up studies as a blunting standard;

[0098] The wear area obtained by instance segmentation is an irregular polygon. The vertex coordinates of the area can be used to further calculate VB max ; First, traverse the vertices of the polygon and obtain the equation of the line segment formed by two points; that is, for the irregular polygon area obtained after segmentation, set the target set T = {t1, t2, ..., t n}, where each target t i is a set of two-dimensional points on the polygonal area. Let point P = (x, y) be a point on the two-dimensional plane, and let a line segment represents the straight line segment connecting points p and q; line segment The length d(p,q) is obtained using the Euclidean distance and its slope is calculated;

[0099] On this basis, the original cutting edge of the tool is determined by filtering out the line segments whose slopes and the number of polygon vertices that do not meet the conditions, such as Figure 5 (a); specifically, traverse all point pairs (p, q) in the target set, calculate the distance and slope, and filter out the line segments that do not meet the conditions; if the number of points passed by the line segment is greater than or equal to the threshold (the threshold set in the present invention is 20), then update the longest distance and line segment, and finally traverse all vertices of the entire polygon to obtain the longest line segment that is the original cutting edge of the tool when it is not worn;

[0100] Then traverse other vertices and calculate their vertical distances to the original cutting edge in turn, that is, given point r = (x3, y3) and line segment Calculate the distance from the point to the line segment; the maximum distance at this time represents the maximum wear width VB max The number of pixel values, such as Figure 5 (b) in the figure; that is, after obtaining the original cutting edge boundary of the tool, it is necessary to determine the point on the wear area farthest from the boundary. The vertical distance between this point and the original cutting edge is the maximum wear width VB of the tool back face. max ;

[0101] By calculating the distance from the point to the line, the maximum wear width VB can be calculated. max The pixel value is further converted based on the actual length and width of the image to obtain VB max The camera software is equipped with basic functions such as measuring area and distance, which can measure the real width and height of the wear image. Because the image resolution is fixed at 3840*2160, VB can be easily obtained after conversion. max The true value of Figure 5 (c) in .

[0102] Example 4

[0103] The multi-domain feature extraction part of multi-sensor information includes two parts: preprocessing and feature extraction; that is, before using the monitoring training sample set to train the wear monitoring model in S3, it also includes: performing noise reduction and feature extraction processing on the monitoring signal to obtain preprocessing features; when training the wear monitoring model, the preprocessing features are used as the input of the wear monitoring model, and the output of the wear monitoring model is the predicted value of the wear value.

[0104] Correspondingly, in S4, the real-time monitoring of the tool wear value is performed using the real-time monitoring signal and the trained wear monitoring model, including: performing noise reduction and feature extraction processing on the real-time monitoring signal to obtain real-time preprocessing features, inputting the real-time preprocessing features into the trained wear monitoring model to obtain the real-time monitoring value of the tool wear value.

[0105] The multiple sensors in S1 include multiple vibration sensors, cutting force sensors, and acoustic emission sensors; the sensors can be installed on the machine tool workbench to collect processing monitoring signals during operation. During preprocessing, since the cutting force, vibration, and acoustic emission data collected by the sensors have a large number of data points, including interference information from factors such as the environment and equipment vibration, these noises will affect the accurate prediction of the tool wear state, so it is necessary to perform noise reduction preprocessing on the multi-sensor raw data; the noise reduction processing of the monitoring signal includes: performing at least one of average filtering, wavelet transform, inverse wavelet transform, and de-idling processing on the monitoring signal; for example:

[0106] First, a filter of size 7 is used to implement average filtering through a sliding window; as shown in formula (1), where S(n+k) represents the value of the original signal at position n+k, K(k) represents the filter, w represents the filter window size, and F(n) represents the filtered signal; it can be seen that the filtered signal is equal to the average value of the signal before filtering after the convolution filter in a window;

[0107] Secondly, the db4 wavelet is used to perform wavelet transform to decompose the signal into details and approximate parts of different scales; then the wavelet coefficients after threshold processing are used to perform inverse wavelet transform to obtain the reconstructed signal; as shown in formula (2), where DWT represents discrete wavelet transform, c j,k represents the approximate partial coefficient, d j,k Represents the detail coefficient, φ j,k (t) represents the scaling function, ψ i,k (t) represents the wavelet function; it is worth noting that only the detail coefficients are processed during threshold processing, and then the approximate coefficients and the detail coefficients after threshold processing are used to reconstruct the signal;

[0108] Finally, since the sensor data collected when the tool is cutting in and out (no-load) has no effect on tool wear and has no characterization capability, the front and back 5% of the data for each pass are eliminated;

[0109]

[0110] The feature extraction process of the monitoring signal includes: extracting multi-domain features from the time domain, frequency domain, and time-frequency domain of the monitoring signal after noise reduction processing; using the Pearson correlation coefficient method to screen out wear-related features (correlation coefficient greater than 0.9) from the multi-domain features; using the KPCA method to map the high-dimensional features of the wear-related features to a low-dimensional space to obtain preprocessing features. In the feature extraction process of the monitoring signal, before using the KPCA method to map the high-dimensional features of the wear-related features to a low-dimensional space, it also includes: performing an exponential filtering operation on the wear-related features, and using an exponential filter with a smoothing coefficient of 0.02 to effectively alleviate the frequent and sudden changes in the wear-related features.

[0111] The monitoring signal is subjected to noise reduction and feature extraction processing to obtain preprocessing features, and also includes: respectively performing noise reduction processing on the monitoring signals obtained by multiple sensors and extracting multi-domain features, and then fusing the multi-domain features corresponding to the monitoring signals obtained by multiple sensors, and then using the Pearson correlation coefficient method based on the fused multi-domain features to screen out wear-strongly correlated features.

[0112] Specifically, when extracting features, the features are extracted as shown in Table 2. In the time domain, the mean, root mean square, and standard deviation are extracted for each signal. The signals of the three directions of cutting force and vibration are extracted and fused with the acoustic emission data to form a 21-dimensional feature. In the frequency domain, the center frequency, frequency variance, and mean square frequency are extracted for each signal according to formula (3). The signals of the three directions of cutting force and vibration are extracted and fused with the acoustic emission data to form a 21-dimensional feature. Taking the 160th pass data of the public data as an example, the frequency spectrum of the three-dimensional force and vibration data obtained by Fourier transform is as follows: Figure 6 As shown in the figure, it can be seen that the frequency of the signal is mainly concentrated in the 1-10Khz frequency band. The spectrum diagram shows that the resolution of each sub-band cannot be less than 518Hz. Therefore, the 1-12.5Khz frequency band is divided into 32 sub-bands in the time-frequency domain, and the energy of the original signal is calculated in each sub-band, resulting in a total of 224 dimensional features.

[0113]

[0114] However, the manually extracted high-dimensional features have more redundant information, which will significantly consume model training time and resources; therefore, the Pearson correlation coefficient method shown in formula (5) is used to select 35-dimensional features with a correlation coefficient greater than 0.9 with tool wear from the 21+21+224-dimensional feature signal as the wear-strong correlation feature;

[0115]

[0116] By observation Figure 7 It is found in the figure after feature extraction and screening that the 0th to 34th feature signals are roughly positively correlated with the 35th wear signal, but the feature signals frequently drop suddenly before and after dimensionality reduction; there may be multiple reasons for this phenomenon; one is that the size of the average filter window setting is unreasonable, another may be that the signal itself contains some rapidly changing events or spikes, or the filter selection is not suitable for processing a specific type of sudden drop in the signal; no matter what the cause is, targeted feature post-processing is required to face this situation, so as not to affect the subsequent model prediction performance; the present invention uses an exponential filtering method with an attenuation factor of 0.02 to eliminate the frequent sudden drops in the signal, such as Figure 8 As shown in (a);

[0117] The post-processed features may be highly correlated with each other, which means they carry repeated information. In order to reduce computational overhead and extract more representative features, the kernel principal component analysis (KPCA) method is used to further map the 35-dimensional high-dimensional features to a low-dimensional space, so that the subsequent model can focus on the truly valuable features in the data, and finally obtain 6-dimensional features, such as Figure 8 As shown in (b); the specific implementation steps of the KPCA feature dimensionality reduction method are shown in Table 3;

[0118] Finally, the samples were divided into units of 3 time steps to construct feature and label data, and the training set and test set were divided into a ratio of 7:3 for use in subsequent monitoring and prediction models.

[0119] Table 2 Multi-domain feature extraction

[0120]

[0121]

[0122] Table 3 KPCA specific implementation steps

[0123]

[0124] Example 5

[0125] The tool wear monitoring model is constructed as a KAN deep learning model; the KAN tool wear real-time monitoring model completes the real-time mapping of multi-sensor multi-domain features to tool wear values; first, the data set is divided into a training set and a test set to evaluate the performance of the model during the training process; secondly, the input features are passed to each neuron in the hidden layer; each hidden layer neuron calculates the univariate function output of its each input variable and sums them, and then performs a nonlinear transformation; thirdly, the outputs of all hidden layer neurons are weighted and summed to obtain the final output; this process realizes the nonlinear combination of high-dimensional inputs through univariate functions to approximate the output of the target function; then, in the training stage, all parameters are randomly initialized, and the difference between the model output and the true label is calculated to obtain the loss function; finally, all parameters of the model are updated with the help of the back-propagation algorithm to obtain the monitoring model with the best performance.

[0126] The tool wear monitoring model shown in formula (6) is used to establish a real-time nonlinear mapping between multi-sensor information and tool wear values, and its output provides a reliable input for the subsequent prediction model; the present invention uses the KAN model to achieve this process;

[0127] y t =M(x t ) (6)

[0128] The KN theorem shown in formula (7) reveals that any multivariable continuous function can be expressed as the superposition of several single-variable continuous functions. This theorem provides a new idea and direction for the deep learning framework. The traditional MLP uses a combination of linear transformation and fixed activation function on the neuron node to complete the nonlinear mapping. KAN directly realizes the mapping between nodes through a learnable nonlinear activation function on the neuron edge, and replaces the linear weight calculation with a parameterized spline single-variable function. Although this innovative method is more complicated than MLP in calculating the relationship between nodes, KAN completely gets rid of the problem of a large number of parameters caused by the fully connected structure of MLP.

[0129]

[0130] n i Indicates i th The number of neurons in the layer, represented by (l,i) th The i-th neuron in the layer, and its activated value is determined by x l,i Indicates; then, between the l and l+1 layers n l n l+1 Among neurons, the activation function connecting (l,i) and (l+1,j) can be expressed as formula (8); therefore, l th The relationship between the output and input can be expressed as formula (9);

[0131] φ l,j,i , l=0,···,L-1, i=1,···,n l , j=1,···,n l+1 (8)

[0132]

[0133] x l The coefficient matrix is ​​represented as Φ l , then the L-layer KAN network can be expressed as formula (10); at this time, by comparing with the L-layer MLP structure of formula (11), it can be seen that KAN replaces the nonlinear mapping composed of the learnable W and the fixed σ of MLP with the learnable Φ; moreover, MLP calculates the linear combination of each neuron in the previous layer on the edge of the neuron and performs nonlinear activation at the node; in contrast, KAN directly performs nonlinear activation on the edge of the neuron and then performs linear combination, thus greatly reducing the computational complexity and saving resources;

[0134]

[0135] Φ is controlled using equation (12-14), where the spline(x) function is parameterized as ci Linear combination with B-spline; Since the parameters are trainable and the transformation of KAN(x) between layers is differentiable, the KAN model can be trained by the back-propagation algorithm;

[0136] φ(x)=w(b(x)+spline(x)) (12)

[0137] b(x)=silu(x / (1+e -x )) (13)

[0138]

[0139] In order to verify the effectiveness of the KAN model, it is compared with other advanced model regression models, as shown in Table 4 and Table 5 respectively; it can be seen that the KAN tool wear real-time monitoring model has smaller error and higher accuracy than other advanced models. C1, C4, and C6 are three existing public data sets. The experimental data set is a data set constructed by experiment based on the wear area segmentation model provided by the present invention to determine the wear value.

[0140] Table 4 Prediction indicators of monitoring model and other advanced models on the public data test set

[0141]

[0142] Table 5 Prediction indicators of monitoring model and other advanced models on the experimental data test set

[0143]

[0144]

[0145] Example 6

[0146] The tool wear prediction model is a Transformer model; the Transformer tool wear prediction model completes the mapping of historical to future tool wear sequences; first, the input features and label data are properly processed and formatted, and divided into training sets and test sets; secondly, the input data is reshaped into a three-dimensional format [batch size, sequence length, features]; this is because the Transformer model is designed based on sequence processing, and it is necessary to clarify the time step and feature dimension of the sequence; then, the model is constructed with the input layer, linear embedding layer, multi-head attention layer, feedforward neural network layer, global average pooling layer, dropout layer and output layer, and the model output is calculated and the loss is calculated with the label; finally, the parameters of the model are iteratively updated through the back-propagation algorithm to obtain the optimal parameter prediction model.

[0147] As shown in formula (15), the tool wear prediction model is used to mine the historical T1 time series data y provided by the monitoring model (t-T1)：t With the future T2 time series data y t.(t+T2) The relationship between them; this process is implemented using the Transformer model;

[0148] y t:(t+T2) =P(y (t-T1):t ) (15)

[0149] First, assume that the model input tensor is Where T represents the sequence length, N represents the batch size, and d represents the input feature dimension. Since the purpose is to achieve the mapping between historical and future wear data, the input feature dimension input = 1. Transformer usually requires high-dimensional input to better capture the complex relationship between features. Therefore, before inputting X into Transformer, it is necessary to first map it to a high-dimensional space through a linear layer. After the change shown in formula (16), its output tensor X0 is

[0150]

[0151] Then X0 is input into the Transformer encoder composed of a multi-head attention mechanism and a feedforward neural network. When performing multi-head attention calculation, for the lth layer, the self-attention calculation of each head h is as shown in formula (17), where WQ, WK, and WV represent the transformation matrices of query, key, and value respectively;

[0152]

[0153] Each Transformer encoder includes a feedforward neural network as shown in formula (18). The output of the lth layer can be expressed as shown in formula (19). After L layers, the output of the encoder can be expressed as shown in formula (20). L In the example, we select the representation of the last time step T, namely X L [-1], and the prediction result can be obtained by regression through the decoder, as shown in (21);

[0154] FFN(X)=ReLU(XW1+b1)W2+b2 (18)

[0155] X l+1 =LayerNorm(FFN(Attention(X l )))+X l (19)

[0156] XL =TransformerEncoder(X0) (20)

[0157]

[0158] In order to verify the effectiveness of the Transformer tool wear prediction model, its performance is first verified in the traditional prediction mode. Its prediction indicators on the public and experimental data test sets are shown in Tables 6 and 7 respectively. It can be seen that the Transformer prediction model has better prediction performance.

[0159] Table 6 Prediction indicators of the prediction model and other advanced models on the public data test set

[0160]

[0161] Table 7 Prediction indicators of the prediction model and other advanced models on the experimental data test set

[0162]

[0163] Example 7

[0164] The advance prediction mode; Fig. 9 The difference between the traditional tool wear prediction and the advanced prediction mode is shown; the traditional tool wear prediction mode relies on historical real data, that is, each prediction requires the monitoring model to provide the historical real wear value; in this mode, the wear value of the historical T1 time point needs to be calculated to be used as the input of the prediction model; therefore, the N predictions of the prediction model require the monitoring model to provide the historical wear data of T1*N time points, which seriously affects the real-time performance of the prediction model;

[0165] In contrast, the advanced prediction mode feeds back the output of the prediction model to the input end and merges it with the historical real wear value data, constantly updating the model input for the next prediction, thereby making continuous predictions; in this mode, theoretically the model can make unlimited recursive predictions based on a set of model inputs, and can make predictions for a longer period of time; however, as the number of predictions increases, the accumulation of errors cannot be ignored, so it is very necessary to use historical tool wear values ​​to correct the model input in a timely manner after a certain number of predictions.

[0166] That is, using the trained wear prediction model to predict the wear value of the tool specifically includes:

[0167] The wear prediction model inputs a monitoring value sequence of tool wear values ​​when predicting for the first time, and outputs a prediction value sequence of tool wear values;

[0168] Afterwards, the wear prediction model inputs a combination of the monitoring value of the tool wear value and the predicted value output from the previous prediction, or a sequence of predicted values ​​output from the previous prediction, and outputs a sequence of predicted values ​​of the tool wear value, forming an advanced prediction mode.

[0169] The monitoring value of the tool wear value is the output value of the wear monitoring model.

[0170] In order to verify the effectiveness of the prediction model in the advance mode, predictions are made on the public and experimental data test sets under different conditions of T1 and T2. The results are shown in Tables 8 and 9, respectively.

[0171] Table 8 Prediction indicators of the prediction model on the public data test set in the advance mode

[0172]

[0173]

[0174] Table 9 Prediction indicators of the prediction model on the experimental data test set in the advance mode

[0175]

[0176] Example 8

[0177] The forward-looking forecasting model also includes:

[0178] Timely correction strategy: comparing the predicted value sequence of tool wear values ​​output by the wear prediction model with the corresponding monitored value sequence of tool wear values ​​to obtain the error between the predicted value and the monitored value;

[0179] If the error is less than the preset threshold, the online prediction is continued in the advance prediction mode; if the error is greater than or equal to the preset threshold, the monitoring value sequence of the tool wear value is input when the wear prediction model makes the next prediction to correct the model input; then the online prediction is continued in the advance prediction mode. During the processing, the input of the wear prediction model can be corrected once in a timely manner at intervals of a preset time or a preset number of cuts.

[0180] Specifically, the timely correction strategy is: in the advance prediction mode, the accumulation of prediction errors leads to an inevitable decrease in prediction accuracy; therefore, a timely correction strategy is proposed, and its principle is shown in Table 10; S i T Indicates that the real historical sequence at the i-th time point prediction is the model input, S i P Y represents the output concatenation of the previous prediction at the i-th time point as the model input; i' represents the prediction result output at the i-th time point, P represents the prediction model, and △ represents the MAE value between the prediction result at the i-th time point and the actual wear sequence; when T1=3, T2=1, the first prediction occurs at the 4th time point, at which time the model is based on the historical sequence S4 T The prediction obtains Y4'; in the future prediction, the prediction can be carried out according to the advance prediction mode, and then the MAE between the previous prediction result and the true value is calculated at each preset time interval or preset number of knife intervals. If it is less than the error threshold δ, then the input of the next prediction will use the previous prediction result fused with historical data as input; on the contrary, if the MAE of the previous prediction is greater than δ, then the next prediction uses the real historical wear data as input, so as to avoid the accumulation of errors in time; among them, the real historical sequence of wear values ​​can be the output value of the wear monitoring model.

[0181] Table 10 Principles of timely correction strategy

[0182]

[0183] Specifically, during the online prediction process, a prediction can be made at regular intervals, and each prediction can be made according to the number of cuts or processing time. Taking the prediction according to the number of cuts as an example: each time the prediction is made, the input of the wear prediction model is a wear value monitoring value sequence corresponding to several cut processings before the current moment, a wear value monitoring value and prediction value fusion sequence, or a wear value prediction value sequence, and the output is a wear value prediction value corresponding to at least one cut processing after the current moment.

[0184] The wear value monitoring value can be obtained based on the wear monitoring model according to the monitoring signal of the sensor collected after the corresponding tool is processed. The wear value monitoring value used for error judgment with the wear value prediction value in timely correction can be obtained in the same way, that is, the number of processing tools corresponding to the wear value prediction value can be obtained based on the wear monitoring model by online collecting the monitoring signal.

[0185] In order to verify the effectiveness of the timely correction strategy in the advance mode, the correction strategy is used to perform predictions on the public and experimental data test sets under different conditions of T1 and T2. The results are shown in Tables 11 and 12, respectively. It can be seen that the use of the timely correction strategy has significantly improved both the error and the accuracy.

[0186] Table 11 Prediction indicators of the prediction model using the correction strategy in the advance mode on the public data test set

[0187]

[0188] Table 12 Prediction indicators of the prediction model using the correction strategy in the advance mode on the experimental data test set

[0189]

[0190] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A real-time monitoring and prediction method for tool wear based on multi-sensor information fusion, characterized in that: include: Offline stage: S1, conduct a machining experiment, collect monitoring signals obtained by multiple sensors and tool wear images obtained by a camera during the machining experiment; S2, using a trained wear area segmentation model to segment the wear area of ​​the tool wear image, and calculating and obtaining a true wear value based on the segmented wear area; S3, using the monitoring signal and the corresponding real wear value to establish a monitoring training sample set, and using the monitoring training sample set to train a wear monitoring model to obtain a trained wear monitoring model; Using the sequence of real wear values ​​to establish a prediction training sample set, using the prediction training sample set to train a wear prediction model, and obtaining a trained wear prediction model; Online stage: S4, during the machining process, using multiple sensors to obtain real-time monitoring signals; using the real-time monitoring signals and the trained wear monitoring model to monitor the wear value of the tool in real time; The trained wear prediction model is used to predict the wear value of the tool.

2. The real-time monitoring and prediction method for tool wear based on multi-sensor information fusion according to claim 1 is characterized in that: Before using the monitoring training sample set to train the wear monitoring model in S3, the following steps are also included: Performing noise reduction and feature extraction processing on the monitoring signal to obtain preprocessing features; when training the wear monitoring model, the preprocessing features are used as inputs of the wear monitoring model, and the output of the wear monitoring model is a predicted value of the wear value; The noise reduction process on the monitoring signal includes: Performing at least one of average filtering, wavelet transform, inverse wavelet transform and dead-load removal processing on the monitoring signal; Performing feature extraction processing on the monitoring signal includes: Multi-domain features are extracted from the time domain, frequency domain, and time-frequency domain of the monitoring signal after noise reduction processing; the wear strong correlation features are screened out from the multi-domain features using the Pearson correlation coefficient method; the high-dimensional features of the wear strong correlation features are mapped to the low-dimensional space using the KPCA method to obtain preprocessing features.

3. The real-time monitoring and prediction method for tool wear based on multi-sensor information fusion according to claim 2 is characterized in that: In the feature extraction process of the monitoring signal, before mapping the high-dimensional features of the wear-strong correlation features to the low-dimensional space using the KPCA method, the method further includes: An exponential filtering operation is performed on the wear-strongly correlated features.

4. The real-time monitoring and prediction method for tool wear based on multi-sensor information fusion according to claim 2 is characterized in that: The monitoring signal is subjected to noise reduction and feature extraction to obtain preprocessing features, further comprising: The monitoring signals acquired by multiple sensors are subjected to noise reduction processing and multi-domain feature extraction respectively, and then the multi-domain features corresponding to the monitoring signals acquired by multiple sensors are fused, and then the Pearson correlation coefficient method is used to screen out the wear-related features based on the fused multi-domain features.

5. The real-time monitoring and prediction method for tool wear based on multi-sensor information fusion according to any one of claims 1 to 4, characterized in that: The wear area segmentation model in S2 is a YOLOv8 segmentation model; by annotating the wear area of ​​the tool wear image obtained in S1, a segmentation sample set is constructed, and the wear area segmentation model is trained using the segmentation sample set.

6. The real-time monitoring and prediction method for tool wear based on multi-sensor information fusion according to any one of claims 1 to 4, characterized in that: In S2, the actual wear value is calculated based on the segmented wear area, specifically including: Obtain pixel coordinates of vertices of the wear area to form a vertex point set; Traversing the vertex point set, obtaining equations of line segments between two vertices; and determining the original cutting edge line segment of the tool by screening the slope of the line segment and the number of vertices passed; Calculate the vertical distances between the remaining points of the vertex point set and the original cutting edge line segment of the tool, and select the value with the largest vertical distance as the number of pixels of the maximum wear width; According to the number of pixels of the maximum wear width and the size of the tool wear image, the actual value of the maximum wear width is obtained as the true wear value.

7. The real-time monitoring and prediction method for tool wear based on multi-sensor information fusion according to any one of claims 1 to 4, characterized in that: The multiple sensors in S1 include multiple vibration sensors, cutting force sensors and acoustic emission sensors; the wear monitoring model in S3 is a KAN deep learning model; the wear prediction model in S3 is a Transformer model.

8. The real-time monitoring and prediction method for tool wear based on multi-sensor information fusion according to any one of claims 1 to 4, characterized in that: The wear prediction model that has been trained is used to predict the wear value of the tool, specifically including: The wear prediction model inputs a monitoring value sequence of tool wear values ​​when predicting for the first time, and outputs a prediction value sequence of tool wear values; Afterwards, the wear prediction model inputs a combination of the monitoring value of the tool wear value and the predicted value output from the previous prediction, or a sequence of predicted values ​​output from the previous prediction, and outputs a sequence of predicted values ​​of the tool wear value, forming an advanced prediction mode. The monitoring value of the tool wear value is the output value of the wear monitoring model.

9. The real-time monitoring and prediction method for tool wear based on multi-sensor information fusion according to claim 8, characterized in that: The forward-looking forecasting model also includes: Timely correction strategy: comparing the predicted value sequence of tool wear values ​​output by the wear prediction model with the corresponding monitored value sequence of tool wear values ​​to obtain the error between the predicted value and the monitored value; If the error is less than the preset threshold, online prediction will continue in the advance prediction mode; if the error is greater than or equal to the preset threshold, when the wear prediction model makes the next prediction, the monitoring value sequence of the tool wear value will be input to correct the model input; thereafter, online prediction will continue in the advance prediction mode.

10. A real-time monitoring and prediction system for tool wear based on multi-sensor information fusion, characterized in that: It comprises a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, it executes the real-time monitoring and prediction method for tool wear based on multi-sensor information fusion as described in any one of claims 1 to 9.

Images