A Method for Predicting the Surface Quality of Machining Based on Multimodal Data

Abstract

The present invention discloses a method for predicting the surface quality of machining based on multimodal data, including: S1. Obtain the process parameters used for machining the current workpiece from a numerically controlled machine tool; S2. Real-time obtain the image of workpiece machining and process it using an image processing model to obtain the surface machining quality grade of the workpiece; S3. Obtain the cutting force and acoustic emission signals of the workpiece during machining and perform preprocessing to obtain the characteristics of the cutting force and acoustic emission signals; S4. Use the surface machining quality grade of the workpiece, the characteristics of the cutting force and acoustic emission signals, and the process parameters as inputs, and the corresponding machining surface quality as the output to train a random forest model; S5. Use the trained random forest model to predict the machining surface quality during each machining. Based on the processing of multimodal data, the present invention can analyze the factors affecting the surface quality of machining from multiple aspects, thereby improving the accuracy of prediction.

Description

Technical Field

[0001] The present invention belongs to the technical field of machining, and particularly relates to a method for predicting the surface quality of machining based on multi-modal data. Background Art

[0002] During the machining process of a numerically controlled machine tool, the surface quality of the workpiece has a direct impact on its service performance. Especially in the machining of parts of precision instruments, the requirements for surface quality are particularly strict, and any deficiency may cause the instrument to malfunction.

[0003] To reduce the production of workpieces with unqualified surface machining quality, the current measures include using machining surface quality prediction technology to pre-evaluate the surface machining quality of the workpiece, so that operators can timely detect potential problems and make corresponding adjustments to the machining parameters. Among them, the offline prediction method of image processing and the online prediction method based on multiple sensors are two relatively common prediction means, both of which can predict the surface machining quality of the workpiece. However, since certain key factors may be missing in the implementation of these two methods, the prediction results may not be accurate enough. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method for predicting the surface quality of machining based on multi-modal data to improve the accuracy of predicting the surface quality of machining through the analysis of multi-modal data in view of the above-mentioned deficiencies of the prior art.

[0005] To achieve the above technical objectives, the technical solution adopted by the present invention is as follows:

[0006] A method for predicting the surface quality of machining based on multi-modal data, comprising:

[0007] S1. Obtain the process parameters used for machining the current workpiece from the numerically controlled machine tool;

[0008] S2. Obtain the image of the workpiece being machined in real time and process it using an image processing model to obtain the surface machining quality grade of the workpiece;

[0009] S3. Obtain the cutting force and the acoustic emission signal of the material during workpiece machining and perform preprocessing to obtain the characteristics of the cutting force and the acoustic emission signal;

[0010] S4. Use the surface machining quality grade of the workpiece, the characteristics of the cutting force and the acoustic emission signal, and the process parameters as inputs, and the corresponding machining surface quality as the output to train a random forest model;

[0011] S5. Use the trained random forest model to predict the machining surface quality during each machining.

[0012] To optimize the above technical solutions, the specific measures also include:

[0013] The process parameters obtained in the above S1 include the spindle speed n, the feed per tooth f z , the cutting depth a p and the cutting width a e .

[0014] The image processing module described in the above S2 is obtained by training the swin-transformer model through the images obtained by using an industrial camera during the machining of the workpiece surface. The specific process includes the following steps:

[0015] S201. Install an industrial camera on the numerically controlled machine tool. During the machining process of each workpiece, take b images of the workpiece surface during machining, and use these b images as a surface image group;

[0016] S202. Classify the surface machining quality of the workpiece, and divide the surface machining quality into a levels; the worst machining quality is defined as level 1, and the best machining quality is defined as level a;

[0017] S203. Select a large number of surface image groups, and define the machining quality level corresponding to each group of images. Use the image group and its corresponding machining quality level as the input and output of the swin-transformer model for training to obtain the image processing model.

[0018] The above S3 specifically includes the following steps:

[0019] S301. Install a force sensor under the workpiece to obtain the cutting force during the machining of the workpiece; install an acoustic emission sensor near the workpiece to obtain the acoustic emission signal of the material during the machining process;

[0020] S302. Intercept the obtained cutting force and acoustic emission signal;

[0021] S303. Perform band-pass filtering on the intercepted data;

[0022] S304. Extract the characteristics of the cutting force and acoustic emission signal after band-pass filtering.

[0023] When the above S302 intercepts the obtained cutting force and acoustic emission signal, select the minimum cutting force F min and the minimum acoustic emission signal I min as the threshold for signal interception, and start the analysis from the first time greater than F min and I min ; and use the condition that the data drops by more than the set value and the distance from the end time of the data is t as the cut-off condition, where t is the average time consumed in the deburring stage.

[0024] The process of the above S303 for performing band-pass filtering on the intercepted data is as follows:

[0025] Assume that the sampling intervals of the force sensor and the acoustic emission sensor are λ1 and λ2 respectively, then the highest frequency of the data is and Set the low-pass frequency of the band-pass filter to f l , and the high-pass frequency to f h . Therefore, the data frequency f(λ) after band-pass filtering is:

[0026]

[0027] where λ takes λ1 or λ2.

[0028] The features of the above S304 include the average cutting force F a and the average acoustic emission signal I a , as well as the variances of the cutting force and the acoustic emission signal F σ , I σ and the root mean squares of the cutting force and the acoustic emission signal F rms , I rms .

[0029] The above S4 specifically includes the following steps:

[0030] One-to-one correspondence between the input and the output is established and divided into a group; each group is numbered, and there are a total of w groups of data; a random interval [w1, w] is set, and v groups of data are randomly selected from it for training during training. The expression of v is as follows: v = rand(w1~w), and it is taken m times. The results of the m times are respectively input into m random forest models for training. Among the m trained random forest models, the result with the largest number of the same test results is taken as the final random forest model.

[0031] The training process of the above single random forest model is as follows:

[0032] 1) Assume that vi groups of data are input into the m i th random forest model, where vi is a random number between [w1, w]. Then the input feature matrix X is a k×vi feature matrix. There are k×vi data in this model. Each data is used as a node to divide the overall data into 2 sub-datasets X1 and X2, where k is the dimension of the input variable;

[0033] 2) Calculate the information entropy after dividing into X1 and X2 and the information entropy before division respectively. The formula for information entropy is as follows:

[0034]

[0035] Among them, p i is the probability of the processing quality under this group of parameters;

[0036] Then the objective function is to maximize the information gain:

[0037]

[0038] Among them, x n is the type of parameter currently in use, and x nj is the division node;

[0039] Treat all k×vi data as division nodes, and find the nodes that satisfy the objective function among them as the first-level root node T 11 ;

[0040] 3) Use the first-level root node T 11 to divide the data into two first-level sub-datasets X 11 , X 12 . For the divided first-level sub-datasets X 11 , X 12 respectively repeat the root node query and sub-dataset division, and obtain the second-level root node T 11 of X 21 and the second-level sub-dataset X 12 and the second-level root node T 22 of the second-level sub-dataset

[0041] 4) Continue to repeat the root node query and sub-dataset division for all the second-level sub-datasets in 3), and keep dividing downwards in the manner of steps 1)-3) until there is a d-level sub-dataset The information gain after division is less than G, then the sub-dataset stops dividing, and other sub-datasets continue to divide until the information gain after division of all levels of sub-datasets is less than the threshold G, then the training of the m i th random forest model is completed.

[0042] The present invention has the following beneficial effects:

[0043] Obtain the current process parameters used for workpiece machining from a numerically controlled machine tool; use an industrial camera to obtain a two-dimensional image of the workpiece machining in real time by means of on-line monitoring; obtain the magnitude of the cutting force during workpiece machining through a force sensor; obtain the acoustic emission signal of the material during the machining process of the machine tool through an acoustic emission sensor; use multi-modal data and its corresponding machining surface quality as input and output, and use an improved random forest algorithm as a training model to obtain the mathematical relationship between each parameter and the machining surface quality, so as to realize the prediction of the machining surface quality. Based on the processing of multi-modal data, the present invention can analyze the factors affecting the machining surface quality from multiple aspects, thereby improving the accuracy of prediction.

[0044] By analyzing the multi-modal data of images, acoustic emission signals and process parameters, the present invention can more comprehensively analyze all factors affecting the machining surface quality of workpieces, and can improve the accuracy of prediction results.

[0045] Aiming at the problem that the random forest algorithm does not set a termination condition, which will lead to too long training time when the sample is too large, the present invention proposes an improved random forest algorithm, which sets an information gain threshold as the termination condition, and can improve the training speed and prediction speed of the model. Brief Description of the Drawings

[0046] Figure 1 It is a flowchart of the method for predicting the machining surface quality of the present invention;

[0047] Figure 2 It is a schematic installation diagram of an industrial camera and a cutting force sensor;

[0048] Figure 3 It is a structural diagram of a swin-transformer;

[0049] Figure 4 It is a flowchart of the prediction algorithm. Detailed Embodiment

[0050] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to 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 used to limit the present invention.

[0051] Although the steps in the present invention are arranged with reference numerals, they are not used to limit the order of the steps. Unless the order of the steps is clearly stated or the execution of a certain step requires other steps as a basis, the relative order of the steps can be adjusted. It can be understood that the term "and / or" used herein refers to and encompasses any and all possible combinations of one or more of the associated listed items.

[0052] Refer to Figures 1-4, a method for predicting the surface quality of machining based on multimodal data according to the present invention is characterized in that the method comprises the following steps:

[0053] S1. Obtain the process parameters used for machining the current workpiece from a numerically controlled machine tool;

[0054] S2. Obtain the image of the workpiece being machined in real time and process it using an image processing model to obtain the surface machining quality grade of the workpiece;

[0055] S3. Obtain the cutting force and acoustic emission signals of the workpiece during machining and perform preprocessing to obtain the characteristics of the cutting force and acoustic emission signals;

[0056] S4. Use the surface machining quality grade of the workpiece, the characteristics of the cutting force and acoustic emission signals, and the process parameters as inputs, and the corresponding machining surface quality as the output to train a random forest model;

[0057] S5. Use the trained random forest model to predict the machining surface quality during each machining.

[0058] In the embodiment, the data sources have a total of three parts: the first part is the combination of process parameters directly obtained from the numerically controlled machine tool; the second part is to obtain the surface image of the workpiece during the machining process through an industrial camera, process the image using the swin-transformer algorithm, classify it according to different machining qualities and assign corresponding grades; the third part is to obtain the cutting force and acoustic emission signals during machining from the force sensor and acoustic emission sensor, intercept the useless data, filter the noise and extract the features of the obtained data to obtain the mean value, variance and root mean square of the cutting force and acoustic emission signals. Then measure the surface machining quality of the workpiece under each set of data, use each set of data and the corresponding surface machining quality as inputs and outputs to train the prediction model, and finally obtain a multimodal mechanical machining surface quality prediction model. When machining the surface of the workpiece each time, the prediction result can be transmitted to the operator to guide the operator to adjust the parameters.

[0059] In S1, the surface machining quality of the workpiece mainly includes surface defects, surface roughness, surface waviness, surface cold work hardening, surface metallographic structure change, surface residual stress, etc. Since the machining process of the machine tool mainly affects the surface defects, surface roughness and surface waviness of the workpiece, the machining surface quality prediction method described in the present invention is mainly used to predict the surface defects, surface roughness and surface waviness of the workpiece.

[0060] Obtain the current process parameters from the numerically controlled machine tool; including the spindle speed n, the feed per tooth f z 、the cutting depth a p and the cutting width a eThese parameters can all be directly obtained from the operation panel of the CNC machine tool, so they will not be elaborated in this article.

[0061] In S2, the obtained images by the industrial camera are used to train the swin-transformer model to obtain an image processing model. For example, install an industrial camera on the machine tool to acquire the images of the workpiece surface during processing, and measure the surface processing quality of the current workpiece. After obtaining a large amount of image data and processing quality data of the workpiece during the processing process, the surface processing quality of the workpiece is classified. For example: the workpieces with obvious defects are classified as level 1, and the ones with the best processing quality are classified as the highest level. After all the surface processing qualities of the workpieces are classified, the image data is used as the input and the processing quality level is used as the output and put into the swin-transformer model for training to obtain an image processing model. The specific process includes the following steps:

[0062] S201, as Figure 2 shown, install an industrial camera obliquely above the workpiece on the CNC machine tool to acquire the images of the workpiece surface during processing. During each workpiece processing process, take b pictures, and take these b pictures as a group.

[0063] S202, classify the processing surface quality of the workpiece, and divide the surface processing quality into a levels; the worst processing quality is defined as level 1, and the best processing quality is defined as level a, then the output interval of the model is

[0064] A = {1, 2,..., a}

[0065] S203, select a large number of surface image groups of workpieces during the CNC machine tool processing, and define the corresponding processing quality level for each group of images; then use the image group and the processing quality level as the input and output of the swin-transformer respectively for training to obtain a reliable image processing model.

[0066] S2031, assume the original pixels of the image are H×W×3, as Figure 3 shown, first pass the image through a PatchPartition module, and use a 4×4 adjacent pixel Patch to flatten in the channel direction, then the dimension of the image at this time becomes Then pass it through a Linear Embeding layer to perform a linear transformation on the channel data of each pixel, making the image become where C is a hyperparameter of the swin-transformer.

[0067] S2032. Input the image processed in step S2021 into a Swin Transformer Block. The Swin Transformer Block is mainly composed of a connection of a W-MSA and a SW-MSA, which can reduce the computational complexity of the transformer algorithm and enable information interaction between different windows, thereby improving the accuracy of classification.

[0068] S2033. Input the image processed in step S2022 into a Patch Merging layer. The Patch Merging layer divides each adjacent pixel into a Patch, then splices the pixels at the same position in the Patch together to obtain 4 feature maps. Then these 4 feature maps are concatenated in the depth direction and passed through a LayerNorm layer. Finally, a linear transformation is performed in the depth direction of the feature map through a fully connected layer, changing the depth of the feature map from C to At this time, the image becomes

[0069] S2034. Then input the image processed in step S2023 into a Swin Transformer Block; then pass it through a Patch Merging layer to change the image to Then pass it through 3 Swin Transformer Blocks; continue to pass it through a Patch Merging layer to change the image to Then pass it through 1 Swin Transformer Block. Finally, pass it through a Layer Norm layer, a global pooling layer, and a fully connected layer to obtain the final output.

[0070] In S3, the cutting force and the acoustic emission signal of the material during the machining process are obtained through a force sensor and an acoustic emission sensor respectively, and data analysis and processing are carried out, mainly including intercepting, denoising, and feature extraction of the data; the specific process includes the following steps:

[0071] S301. As Figure 2 Install a force sensor under the workpiece to be machined to measure the magnitude of the cutting force during the workpiece machining process; install an acoustic emission sensor near the workpiece to measure the acoustic emission signal of the material during the machining process.

[0072] S302. Since there is no-load data before machining and the deburring step after machining has no effect on the cutting force, it is necessary to intercept the obtained cutting force data and acoustic emission signal.

[0073] In the no-load stage, both the cutting force and the acoustic emission signal are relatively small, showing an obvious gap compared with the working stage. Therefore, the minimum cutting force F min and the minimum acoustic emission signal I min are selected as the thresholds for signal interception. During analysis, it will start from the first time greater than F min and I min .

[0074] In the deburring stage, since the time of the whole process is relatively fixed, the average time t consumed in the deburring stage is taken; and the cutting force and the acoustic emission signal in the deburring stage are lower than those in the working stage. Therefore, the condition for termination is that the data shows an obvious decline and is at a distance of t from the end time of the data.

[0075] S303. Since the noise in the machining process will affect the accuracy of the data, the intercepted data is subjected to band-pass filtering. Assuming that the sampling intervals of the force sensor and the acoustic emission sensor are λ1 and λ2 respectively, the highest frequencies of the data are and The low-pass frequency of the band-pass filter is set to f l , and the high-pass frequency is f h . Therefore, the frequency expression f(λ) of the filtered data is:

[0076]

[0077] S304. Extract the characteristics of the processed cutting force and acoustic emission signal, including the mean value F a of the cutting force and the mean value I a of the acoustic emission signal, as well as the variances F σ , I σ of the cutting force and the acoustic emission signal, and the root mean squares F rms , I rms of the cutting force and the acoustic emission signal. Specifically, the following formulas are as follows:

[0078]

[0079]

[0080]

[0081]

[0082]

[0083]

[0084] Among them, N is the number of processed data collected during the processing, i is the serial number of the data, and i = 1, 2,..., N.

[0085] In S4, the above-obtained data (as Figure 4 shown, from the numerical control machine tool, the swin-transformer model, and the data processed by the sensor) are respectively used as the input, and the final surface machining quality is used as the output to train the improved random forest prediction model; the specific process includes the following steps:

[0086] (1) Input the image information during the machining process of the current workpiece into the swin-transformer model in step S203 to obtain the surface machining quality grade of the current workpiece; then obtain each feature of the current force sensor and acoustic emission sensor after processing through S304.

[0087] (2) The process parameters obtained from the numerical control machine tool: spindle speed n, feed per tooth f z , cutting depth a p and cutting width a e ; the surface machining quality grade A obtained in step S203, the average cutting force F obtained in S304 a and the average acoustic emission signal I a , the variances of the cutting force and acoustic emission signal F σ , I σ and the root mean squares of the cutting force and acoustic emission signal F rms , I rms are used as the input, and the surface machining quality of the current workpiece is used as the output.

[0088] (3) The above inputs and outputs are corresponded one by one and divided into a group; each group is numbered, and there are a total of w groups of data. Then set a random interval [w1, w], where w1 ≤ w, and randomly select v groups of data for training during training. The expression of v is as follows:

[0089] v = rand(w1~w)

[0090] A total of m times are taken, and the results of the m times are respectively input into m models for training.

[0091] (4) The training process of the model is as follows:

[0092] 1) Assume that the vi group of data is input into the m i th model, where vi is a random number between [w1, w], then the input feature matrix X is:

[0093]

[0094] Then there are 11×vi data in this model. Each data is regarded as a node, and then the overall data is divided into two sub-datasets X1 and X2. Taking the rotational speed n as an example, assume that n j is selected as the node, then the dataset is divided into:

[0095] X1 = {n l |n i ≤n j , i∈[1, vi]}

[0096] X2 = {n h |n i >n j , i∈[1, vi]}

[0097] where i is the row number of the overall data. Then the two sub-datasets X1 and X2 obtained by taking n j as the node are:

[0098]

[0099]

[0100] where l1, l2, …, lk and h1, h2, …, h(vi - k) are the row numbers of the newly divided dataset.

[0101] 2) Calculate the information entropy after dividing into X1 and X2 and the information entropy before division respectively. If the information entropy is larger, it means the information chaos degree is higher. The formula of information entropy is as follows:

[0102]

[0103] where p i is the probability of the processing quality under this set of parameters. Subtracting the information entropy before and after division can obtain the information gain formula; the larger the information gain, the better the effect of the segmentation process. In order to make the information gain maximum, the objective function should be:

[0104]

[0105] where x n is the type of parameter currently used, and x nj is the division node. Then repeat the above steps until all 11×vi data are regarded as division nodes, and find the node with the maximum information gain as the first-level root node T 11 .

[0106] 3) Since the information gain will be very small when two sub - datasets are relatively close, and relatively close surface machining qualities can be classified into one category, a threshold of information gain G is set to improve the running speed of the model. Using the first - level root node T 11 Split the data into two first - level sub - datasets X 11 、X 12 , and respectively repeat the root - node query and sub - dataset division on the split first - level sub - datasets X 11 、X 12 to obtain the second - level root node T 11 of X 21 and the second - level sub - datasets X 12 and the second - level root node T 22 of the second - level sub - datasets where the first digit of the subscript is the level of the sub - dataset, the second digit of the subscript is the sub - dataset serial number, and the superscript is the serial number of the upper - level sub - dataset.

[0107] 4) Continue to repeat the root - node query and sub - dataset division for all the second - level sub - datasets in step 3), and keep dividing downwards according to the methods in steps 1) - 3) until the information gain of the d - level sub - datasets after division is less than G, then the sub - dataset stops dividing, and other sub - datasets continue to divide. Until the information gain of all levels of sub - datasets after division is less than the threshold G, the model training is completed.

[0108] (5) Input the parameters corresponding to the workpiece to be predicted into the m trained models, and take the result with the largest same - number of test results as the final output.

[0109] In S5, during each machining, predict the surface machining quality, transfer the prediction result to the CNC machine tool control panel to guide the operator to make adjustments, and the operator can adjust the parameters used according to the prediction result.

[0110] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above - mentioned exemplary embodiments, and can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention. Therefore, in any aspect, the embodiments should be regarded as exemplary and non - restrictive. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be encompassed by the present invention. Any reference signs in the claims should not be regarded as limiting the claimed rights.

[0111] In addition, it should be understood that although this specification is described according to embodiments, not every embodiment only contains an independent technical solution. This narrative way of the specification is only for clarity. Those skilled in the art should regard 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 method for predicting the surface quality of machining based on multi-modal data, characterized in that, Including: S1. Obtain the process parameters used for the current workpiece machining from the numerical control machine tool; S2. Obtain the image of the workpiece machining in real time and process it using an image processing model to obtain the surface machining quality grade of the workpiece; S3. Obtain the cutting force and the acoustic emission signal of the material during workpiece machining and perform preprocessing to obtain the characteristics of the cutting force and the acoustic emission signal; S4. Use the surface machining quality grade of the workpiece, the characteristics of the cutting force and the acoustic emission signal, and the process parameters as inputs, and the corresponding machining surface quality as the output to train a random forest model; S5. Use the trained random forest model to predict the machining surface quality during each machining; The image processing model in S2 is obtained by training the swin-transformer model through the images obtained by using an industrial camera during the surface machining of the workpiece. The specific process includes the following steps: S201. Install an industrial camera on the numerical control machine tool. During each workpiece machining process, take b images of the workpiece surface machining, and use these b images as a surface image group; S202. Classify the machining surface quality of the workpiece, and divide the surface machining quality into a levels; the worst machining quality is defined as level 1, and the best machining quality is defined as level a; S203. Select a large number of surface image groups, and define the machining quality grade corresponding to each group of images. Use the image group and its corresponding machining quality grade as the input and output of the swin-transformer model respectively for training to obtain the image processing model; The specific content of S203 includes: S2031. Assume that the original pixels of the image are . First, pass the image through a Patch Partition module. Flatten a 4×4 Patch of adjacent pixels in the channel direction. At this time, the dimension of the image becomes . Then, pass through a Linear Embeding layer to perform a linear transformation on the channel data of each pixel, making the image become . where C is the hyperparameter of the swin-transformer; S2032. Input the image processed in step S2031 into a Swin Transformer Block, and the Swin Transformer Block is composed of a W-MSA and a SW-MSA connected; S2033. Input the image processed in step S2032 into a Patch Merging layer. The Patch Merging layer divides every adjacent pixel into a patch, then stitches together the pixels at the same position in the patch to obtain 4 feature maps, concatenates these 4 feature maps in the depth direction, then passes through a LayerNorm layer, and finally passes through a fully connected layer to perform a linear transformation in the depth direction of the feature map, changing the depth of the feature map from C to . At this time, the image becomes ; S2034. Input the image processed in step S2033 into a Swin Transformer Block, and then use a Patch Merging layer to transform the image into , and then pass it through 3 Swin Transformer Blocks; continue to use a Patch Merging layer to transform the image into , then pass it through 1 Swin Transformer Block, and finally pass it through a Layer Norm layer, a global pooling layer, and a fully connected layer to obtain the final output; The specific content of S4 includes the following steps: Correspond the input to the output one by one and divide them into a group; number each group, and there are w groups of data in total; set a random interval , and randomly select v groups of data for training during training, v and the expression is as follows: , take a total of m times, input the results of m times into m random forest models for training respectively. Among the m trained random forest models, take the result with the largest number of identical test results as the final random forest model; The training process of a single random forest model is: 1) Assume that in the th random forest model, groups of data are input, where is a random number between, then the input feature matrix X is a feature matrix of k × vi . Then in this model, there are k × vi data. Taking each data as a node respectively, the overall data is divided into two sub-datasets and , where k is the dimension of the input variable; 2) Calculate the information entropy after being divided into and respectively, and the information entropy before division. The formula for information entropy is as follows: ； Among them, is the probability of the machining quality under this set of parameters; Then the objective function is to maximize the information gain: ； Among them, is the parameter type currently in use, is the division node; Take all k × vi data as division nodes, and find the nodes that satisfy the objective function among them as the first-level root nodes ; 3) Use the first-level root node Divide the data into two first-level sub-datasets , , and for the divided first-level sub-datasets , Repeat the root node query and sub-dataset division respectively to obtain the second-level root node and the second-level sub-datasets , ; the second-level root node and the second-level sub-datasets , ; 4) Continuously repeat the root node query and the division of the sub-datasets for all the secondary sub-datasets in 3), and keep dividing downward in the manner of steps 1)-3) until there are d-level sub-datasets If the information gain after division is less than G, then the division of the sub-dataset stops, and the other sub-datasets continue to be divided until the information gain after the division of all levels of sub-datasets is less than the threshold G, then the training of the first random forest model is completed.

2. The method for predicting the surface quality of machining based on multi-modal data according to claim 1, characterized in that The process parameters obtained by S1 include the spindle speed n, the feed per tooth f z , the cutting depth a p and the cutting width a e .

3. A method for predicting the surface quality of machining based on multi-modal data according to claim 1, characterized in that, The specific content of S3 includes the following steps: S301. Install a force sensor under the workpiece to obtain the cutting force during the workpiece machining process; install an acoustic emission sensor near the workpiece to obtain the acoustic emission signal of the material during the machining process; S302. Intercept the obtained cutting force and acoustic emission signal; S303. Perform band-pass filtering on the intercepted data; S304. Extract the characteristics of the cutting force and the acoustic emission signal after the band-pass filtering process.

4. A method for predicting the surface quality of machining based on multimodal data according to claim 3, characterized in that, When intercepting the obtained cutting force and acoustic emission signals in S302, the minimum cutting force F min and the minimum acoustic emission signal I min are selected as the thresholds for signal interception, and the analysis starts from the first time greater than F min and I min ; and the condition for termination is that the data drops by more than the set value and the distance from the end time of the data is t, where t is the average time consumed in the deburring stage.

5. A method for predicting the surface quality of machining based on multi-modal data according to claim 3, characterized in that The process of performing band-pass filtering on the intercepted data in S303 is: Assume that the sampling intervals of the force sensor and the acoustic emission sensor are λ1 and λ2 respectively, then the highest frequency of the data is and ; Set the low-pass frequency of the band-pass filter to be f l , and the high-pass frequency to be f h . Therefore, the data frequency f(λ) after band-pass filtering is: ； where λ takes λ1 or λ2.

6. A method for predicting the surface quality of machining based on multi-modal data according to claim 3, characterized in that, The characteristics of the S304 include the mean value of the cutting force the mean value of the acoustic emission signal , and the variances of the cutting force and the acoustic emission signal 、 and the root mean square of the cutting force and the acoustic emission signal 。

Images