Polishing operation control method and polishing operation control system
By obtaining three-dimensional information on the workpiece surface and generating a polishing solution using a roughness prediction model, the problem of inaccurate control of polishing factors in the prior art is solved, and the polishing accuracy and product yield are improved.
Patent Information
- Application Number
- CN202510095805.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-22
- Publication Date
- 2025-05-02
AI Technical Summary
The prior art is difficult to accurately control polishing factors in polishing operations, resulting in the smoothness and size of the workpiece not meeting the requirements, and reduces the polishing efficiency and product yield.
By obtaining three-dimensional information on the workpiece surface, a polishing area is generated, and a polishing scheme is generated based on a pre-trained roughness prediction model, the polishing data is collected in real time to control the polishing operation.
Accurate control of polishing factors is achieved, polishing accuracy and product yield are improved, and the number of repeated polishing and testing is reduced.
Smart Images
Figure CN119910507A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the technical field of polishing, and in particular to a polishing operation control method and a polishing operation control system. Background Art
[0002] Polishing is the process of polishing the surface of a workpiece. The most commonly used polishing method is to polish the workpiece with a polishing head. During the polishing operation, the polishing liquid is sprayed onto the area to be polished. The abrasive particles in the polishing liquid will cut the surface of the workpiece under the drive of the polishing head, thereby polishing the workpiece to a predetermined smoothness.
[0003] In actual production, there are workpieces that require high polishing accuracy and require different smoothness on different surfaces of the finished workpiece, and different surfaces of the finished workpiece are required to present different smoothness so that different surfaces of the finished workpiece have different functional characteristics. The smoothness of the finished workpiece is affected by many polishing factors, such as polishing time, polishing head speed, relative position of the polishing head and polishing surface, and the size of abrasive particles in the polishing liquid.
[0004] Due to the lack of overall planning for polishing operations, it is difficult to accurately control various polishing factors during polishing operations in the prior art, and it is often necessary to repeatedly polish and test the smoothness of the workpiece to achieve the expected smoothness requirements. This operation mode of the prior art reduces the polishing efficiency on the one hand, and on the other hand, since the workpiece needs to be repeatedly polished and tested for smoothness, although the finished workpiece can meet the expected smoothness requirements, the probability that the size of the finished workpiece cannot meet the expected requirements is greatly increased, resulting in a decrease in the yield rate of the product. Summary of the invention
[0005] The content of this application is used to introduce concepts in a brief form, which will be described in detail in the detailed implementation section below. The content of this application is not intended to identify the key features or essential features of the technical solution claimed for protection, nor is it intended to limit the scope of the technical solution claimed for protection.
[0006] As a first aspect of the present application, in order to solve the technical problems mentioned in the above background technology section, some embodiments of the present application provide a polishing operation control method, comprising the following steps:
[0007] Step 1: Obtain three-dimensional information of the workpiece surface and generate a polishing area on the workpiece surface;
[0008] Step 2: Obtain polishing requirements for each polishing area, and generate a polishing plan based on the polishing requirements;
[0009] Step 3: Polish the workpiece based on the polishing plan and collect polishing data in real time;
[0010] The polishing data includes the strength of the polishing head, the rotation speed of the polishing head, the curvature of the polishing area and the abrasive concentration of the polishing liquid;
[0011] Step 4: Setting a prediction device, which has a built-in roughness prediction model;
[0012] Step 5: Input the polishing plan into the roughness prediction model to generate predicted polishing data, and control the polishing operation based on the predicted polishing data.
[0013] The technical solution of the present application generates an overall polishing plan for the workpiece before the polishing operation begins. Based on the setting of the pre-trained roughness prediction model, the polishing level under different polishing data can be predicted. Based on the setting of the roughness prediction model, estimated polishing data can be generated, and the workpiece is polished according to the estimated polishing data, which can accurately control various polishing elements and improve polishing accuracy and product yield. The technical solution of the present application solves the problem that the prior art requires repeated polishing and testing of the workpiece, greatly improves the polishing efficiency, reduces the probability that the workpiece size does not meet the requirements, and improves the yield rate.
[0014] Since the three-dimensional shape of the workpiece is ever-changing, collecting the three-dimensional information of the entire workpiece is not only time-consuming, but also when the workpiece is too detailed, the amount of three-dimensional information obtained will be extremely large. To this end, this application proposes an innovative technical solution, the specific steps are as follows:
[0015] Step 1 includes the following steps:
[0016] Step 11: Identify all areas on the workpiece that need to be polished and use them as polishing areas.
[0017] Step 12: Collect three-dimensional information of the surface of the polishing area and calculate the curvature of the polishing area.
[0018] In the technical solution proposed in this application, it is not required to obtain the overall three-dimensional information of the workpiece, but only focus on the polishing area involved in the actual polishing operation to obtain its three-dimensional surface information relative to the polishing head. In this way, the amount of information in this solution is more concise, and information redundancy is effectively reduced.
[0019] During the polishing process, the surface of the workpiece may be damaged to a certain extent, so it is necessary to minimize polishing errors and improve polishing accuracy to avoid significant changes in the size of the workpiece after multiple polishing. To this end, this application proposes the following technical solutions:
[0020] Further, step 2 includes the following steps:
[0021] Step 21: Calculate the difference between the initial polishing level and the target polishing level of each polishing area to obtain a mark value of each polishing area;
[0022] Step 22: sort all polishing areas according to the mark value from small to large to form a polishing area sequence;
[0023] Step 23: Polish the polishing area with a small mark value first, and then polish the polishing area with a large mark value to generate a polishing plan.
[0024] During the polishing process, the larger the mark value, the more workpiece surface material needs to be cut, which is more likely to cause the workpiece size to change. Therefore, the technical solution provided by this application fully considers the mark value of each polishing area when formulating the polishing plan. By first polishing the polishing area with a small mark value, and then using the roughness prediction model to correct it, the accuracy of subsequent polishing is improved, thereby ensuring that a high accuracy can be achieved when polishing the area with a large mark value, reducing the number of repeated polishing that may occur in such areas with large mark values, and reducing the risk of reduced dimensional accuracy of the workpiece after multiple polishing.
[0025] The polishing roughness of the workpiece surface is affected by many factors. The sample data used in the training of the roughness prediction model is often quite different from the actual situation, which may lead to inaccurate prediction results.
[0026] Further, step 3 includes the following steps:
[0027] Step 31: Based on the polishing plan, determine the execution order of each polishing area on the workpiece, and polish each polishing area according to the execution order;
[0028] Step 32: After polishing of each polishing area is completed, the polishing level of the polishing area and the polishing data of this polishing are obtained, and the obtained polishing level and polishing data of this polishing are input into the prediction device.
[0029] The prediction device includes:
[0030] Prediction module, with built-in roughness prediction model for predicting roughness;
[0031] The updating module receives the roughness level of the polishing area after polishing and the polishing data of this polishing, and inputs them into the prediction module to feedback and update the roughness prediction model.
[0032] In the technical solution proposed in the present application, when polishing the workpiece, each area is polished separately, so after each polishing is completed, the roughness prediction model can be updated in time according to the feedback of the polishing data.
[0033] When predicting the roughness of the workpiece surface, a lot of information needs to be considered, such as the contact surface between the polishing head and the workpiece, the size of the abrasive particles in the polishing liquid, the elasticity of the abrasive particles, etc. By building a corresponding model with this information, it is easier to build an accurate prediction model. However, this prediction model takes a lot of time to calculate, which reduces the polishing efficiency. To this end, the present application provides the following technical solutions:
[0034] The following steps are performed when training the roughness prediction model:
[0035] S1: construct a teacher model and train the teacher model with complete modeling data, wherein the complete modeling data includes polishing data, polishing trajectory data, free abrasive particle impact data, effective abrasive particle impact data, and abrasive particle and workpiece elasticity data;
[0036] S2: Input the training samples into the teacher model and the roughness prediction model respectively, use the loss between the roughness prediction model and the teacher model as the distillation loss, and update the built-in parameters of the roughness prediction model;
[0037] Among them, the roughness in the training sample is the prediction information, and the polishing data is the input information.
[0038] In the technical solution provided by the present application, a teacher model is pre-constructed when training the roughness model. The teacher model will be trained with more complete and complex polishing trajectory data, free abrasive impact data, effective abrasive impact data, and abrasive and workpiece elasticity data, so that the prediction accuracy of the teacher model is higher. Then, the teacher model is used to perform distillation learning on the roughness prediction model. In this way, although the model structure of the roughness prediction model is simple, by learning the teacher model's ability to understand the training samples, the roughness prediction model can obtain the teacher model's ability to understand the training samples, thereby ensuring the prediction accuracy.
[0039] In the polishing operation, the polishing data affects the roughness of the polishing surface, rather than the roughness of the polishing surface affecting the polishing data. Therefore, when constructing a roughness prediction model, the model can only reveal the influence of polishing data on roughness, but cannot infer the corresponding polishing data based on the known roughness. To this end, the present application proposes the following technical solutions:
[0040] The prediction device includes an information integration module;
[0041] The information integration module is connected to the prediction module signal, and is used to obtain the corresponding relationship between the polishing data and the roughness level in the prediction module, and screen out the predicted polishing data based on the input roughness level.
[0042] The technical solution provided in this application, by setting up an information integration module, enables the pre-trained roughness prediction model to accurately guide the polishing operation.
[0043] In practice, when the information integration module screens out the expected polishing data, it needs to traverse all types of polishing data, which takes a lot of time, thereby reducing the polishing efficiency. To this end, the present application provides the following technical solutions:
[0044] The information integration module selects the expected polishing data based on the following steps:
[0045] S01: Generate a polishing information pre-selection matrix based on the input target polishing level;
[0046]
[0047] The rows in the polishing information pre-selection matrix K represent the values of each polishing data, and the columns in the polishing information pre-selection matrix K represent the types of polishing data; n represents the total number of types of polishing data, k 1,1 Indicates the first value in the first type of polishing data; k 1,… Indicates the last value in the first type of polishing data; k n,… Indicates the last value in the nth polishing data;
[0048] S02: configure the objective function f(x);
[0049] f(x) = Q1 + Q2;
[0050] Among them, x is the input of the objective function f(x), Q1 is the polishing data change score, which is negatively correlated with the polishing data change rate; Q2 is the polishing data close score;
[0051]
[0052] Among them, w n Represents the weight score of the nth polishing data, r n Indicates the change rate of the nth polishing data under the current value;
[0053]
[0054] Among them, a represents the number of iterations index, Y a represents the polishing level predicted by the selected polishing data input into the roughness prediction model at the ath iteration, and Y0 represents the target polishing level;
[0055] S03: Preselect matrix K based on polishing information to construct feasible solution matrix D, set the maximum number of iterations, number of particles, particle initial velocity position x ij ;
[0056]
[0057] Among them, x ij is the value of the jth dimension of the ith particle, and Respectively represent x ij The maximum and minimum values of, i∈{1,2,…C}, j∈{1,2,…h}, rand(0,1) is used to generate any random number between 0 and 1, and h represents the dimension of the feasible solution matrix D;
[0058] S04: Update the particle position based on the maximum objective function, repeat the above steps, and perform m iterations continuously until the maximum number of iterations is reached or Q2 reaches the threshold.
[0059] In the technical solution provided in the present application, a polishing data traversal matrix is constructed and updated and iterated using a particle swarm algorithm, so that the closest polishing data can be screened out in a short time.
[0060] As a second aspect of the present application, a polishing operation control system is provided, comprising:
[0061] An information input module, used to input an expected polishing grade and polishing plan;
[0062] A prediction device, wherein the prediction device has a built-in roughness prediction model;
[0063] A control module controls the polishing level based on the polishing operation control method. BRIEF DESCRIPTION OF THE DRAWINGS
[0064] The drawings constituting a part of this application are used to provide a further understanding of this application, so that other features, purposes and advantages of this application become more obvious. The illustrative embodiment drawings and their descriptions of this application are used to explain this application and do not constitute an improper limitation on this application.
[0065] In addition, throughout the drawings, the same or similar reference numerals represent the same or similar elements. It should be understood that the drawings are schematic and that the components and elements are not necessarily drawn to scale.
[0066] In the attached picture:
[0067] Figure 1 The figure is a flow chart of the polishing operation control method.
[0068] Figure 2 Schematic diagram of the structure of the roughness prediction model and the teacher model. DETAILED DESCRIPTION
[0069] Embodiments of the present application will be described in more detail below with reference to the accompanying drawings. Although certain embodiments of the present application are shown in the accompanying drawings, it should be understood that the present application can be implemented in various forms and should not be construed as being limited to the embodiments set forth herein. On the contrary, these embodiments are provided to provide a more thorough and complete understanding of the present application. It should be understood that the drawings and embodiments of the present application are only for exemplary purposes and are not intended to limit the scope of protection of the present application.
[0070] It should also be noted that, for ease of description, only the parts related to the invention are shown in the drawings. In the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other.
[0071] The present application will be described in detail below with reference to the accompanying drawings and in combination with embodiments.
[0072] The special polishing process used in this application is grinding and polishing. Specifically, the polishing head is usually circular in design and rotates under power drive. During the polishing process, the polishing head is in close contact with the workpiece, and then the workpiece is finely polished through the selective movement of the polishing head. During the polishing operation, the polishing position is continuously sprayed with a polishing liquid containing granular abrasive particles. Under the action of the polishing head, these abrasive particles can effectively cut the surface of the workpiece, so that the surface of the workpiece gradually becomes smooth and flat. The polishing head in this scheme adopts an airbag design, that is, by adjusting the air pressure, the hardness, elasticity and pressure of the polishing head during polishing can be flexibly changed. This polishing process belongs to the scope of the prior art, so the specific details of how to use this process for polishing will not be repeated in this article.
[0073] Example 1: Reference Figure 1 , the polishing operation control method includes the following steps:
[0074] Step 1: Obtain three-dimensional information of the workpiece surface and generate a polishing area on the workpiece surface;
[0075] Step 1 includes the following steps:
[0076] Step 11: Identify all areas on the workpiece that need to be polished and use them as polishing areas.
[0077] Not all areas on the workpiece need to be polished. Instead, according to specific polishing requirements, the areas that need to be polished are determined and set as polishing areas. When demarcating the polishing area, the area with the same relative position as the polishing head is divided into one polishing area. Specifically, when demarcating the polishing area, it is not important whether the areas are connected, but whether the relative position of the polishing head and these areas remains consistent during the polishing process. For example, for a regular hexagonal workpiece, the relative position of each face with the polishing head remains unchanged during polishing, and these six faces are regarded as one polishing area. On the contrary, for another workpiece, if the curvature of each face is different, then during polishing, the relative position of the polishing head and each face will also be different. Therefore, in this case, each face needs to be separately demarcated as a polishing area.
[0078] The above is the principle of dividing the polishing area in this solution. The reason for adopting such a division method is to ensure that the relative position relationship between each position and the polishing head in the same polishing area is consistent, so as to ensure the uniformity and consistency of the polishing effect, while avoiding too many polishing areas.
[0079] Step 12: Collect three-dimensional information of the surface of the polishing area and calculate the curvature of the polishing area.
[0080] The calculation method of the polishing area surface area is prior art and will not be repeated here. The reason for calculating the curvature is that the curvature of the polishing surface indicates the degree of bending, which will affect the contact area and contact pressure between the workpiece and the polishing head during polishing, thereby affecting the polishing work.
[0081] Step 2: Obtain polishing requirements for each polishing area, and generate a polishing plan based on the polishing requirements.
[0082] Step 2 includes the following steps:
[0083] Step 21: Calculate the difference between the initial polishing level and the target polishing level of each polishing area to obtain a mark value of each polishing area.
[0084] In some optical instruments, the smoothness requirements of various surfaces of parts (such as prisms) are not consistent. In order to meet this requirement, different surfaces of the parts need to be polished to different degrees. In this scheme, the difference in roughness is described by the polishing grade. The higher the polishing grade, the smoother the workpiece surface, and vice versa.
[0085] In order to clarify the specific polishing requirements of each polishing area, the concept of marking value is introduced. For example, when a workpiece is produced, the initial polishing level of a polishing area is 1. According to the requirements, the area needs to be processed to polishing level 10, so the marking value of the polishing area is 9. Therefore, the larger the marking value, the greater the polishing difficulty.
[0086] The initial polishing level in this solution refers to the polishing level of the workpiece after it is produced and has not yet been polished. It can also be understood as the original state before polishing. The target polishing level is the desired polishing level set according to actual needs. The initial polishing level in this solution is the original polishing level after the workpiece is produced, or the polishing level before polishing. The target polishing level is the required polishing level set according to demand.
[0087] Step 22: Sort all polishing areas according to the mark value from small to large to form a polishing area sequence.
[0088] After calculating the mark value of each polishing area, a polishing area sequence can be formed.
[0089] Step 23: Polish the polishing area with a small mark value first, and then polish the polishing area with a large mark value to generate a polishing plan.
[0090] Since different target polishing requirements require different polishing processes, when dividing the polishing area, it is also necessary to divide it according to the target polishing level. If in the previous division scheme, the target polishing levels of two different areas in the same polishing area are different, then this polishing area needs to be divided into two polishing areas based on the range of the target polishing level.
[0091] The polishing scheme mainly refers to the execution order of each polishing area when the workpiece is polished. As for how to switch the polishing area during the polishing work, this belongs to the prior art and will not be repeated here.
[0092] Step 3: Polish the workpiece based on the polishing scheme and collect polishing data in real time; the polishing data includes polishing head strength, polishing head rotation speed, polishing area curvature and polishing liquid abrasive concentration.
[0093] Step 3 includes the following steps:
[0094] Step 31: Based on the polishing plan, determine the execution order of each polishing area on the workpiece, and polish each polishing area according to the execution order;
[0095] Step 32: After polishing of each polishing area is completed, the polishing level of the polishing area and the polishing data of this polishing are obtained, and the obtained polishing level and polishing data of this polishing are input into the prediction device.
[0096] The core of this solution is to build a roughness prediction model. This model uses a neural network architecture and is an empirical model that can deeply explore the intrinsic correlation between the polishing level (i.e., roughness) and the polishing data. By inputting polishing data into this roughness prediction model, the polishing level that can be achieved after polishing with this data can be predicted. Therefore, in order to improve the prediction accuracy, this solution requires that after the polishing operation is completed in each polishing area, the system collects the polishing data and the corresponding polishing level of the time for training or updating the roughness prediction model.
[0097] Step 3 elaborates on the need to perform polishing operations in sequence according to the polishing area sequence set in the polishing plan, and after polishing of each polishing area is completed, necessary information needs to be collected for subsequent analysis.
[0098] Step 4: Setting a prediction device, which has a built-in roughness prediction model;
[0099] The prediction device includes:
[0100] Prediction module, with built-in roughness prediction model for predicting roughness;
[0101] The updating module receives the roughness level of the polishing area after polishing and the polishing data of this polishing, and inputs them into the prediction module to feedback and update the roughness prediction model.
[0102] Among them, the prediction model is used to store and calculate the roughness prediction model, and the update module will feedback and update the roughness model according to the roughness level after polishing and the polishing data of this polishing.
[0103] The prediction module is a separate processor. The update module is also a separate processor, which is used to control the roughness level of the polishing area after polishing is completed and input the polishing data of this polishing into the prediction module.
[0104] In the conventional training process of the neural network model, the model is usually fixed after the training phase is completed, and its built-in parameters are no longer updated when used. However, this solution adopts a different strategy and sets up an update module to provide real-time feedback and update of the roughness prediction model. Specifically, this update method is consistent with the update logic of the roughness prediction model in the initial training phase. It only switches the working state of the roughness prediction model from the prediction state to the training state to achieve continuous optimization and adjustment of the model.
[0105] The roughness prediction model in this solution is the core model for realizing polishing grade control. In this embodiment, the roughness prediction model is an LSTM model, which includes multiple LSTM units, each of which includes:
[0106] Input Gate: controls the extent to which new information enters the cell.
[0107] Forget Gate: controls the extent to which old information is forgotten.
[0108] Cell State: Stores long-term information and is updated through the forget gate and input gate.
[0109] Output Gate: controls the output of cell state.
[0110] Each LSTM unit receives the current input data, the hidden state and cell state of the previous time step, outputs the new hidden state and cell state after a series of calculations, and finally outputs the polishing level.
[0111] Under the premise of known model structure, the training method of the roughness prediction model belongs to the scope of existing technology and will not be elaborated in detail here.
[0112] For ease of understanding, the following will introduce the various components of the polishing data in detail: Polishing data includes polishing head strength, polishing head rotation speed, polishing area curvature and polishing liquid abrasive concentration. In this solution, the polishing head adopts an airbag design, so the polishing head strength is closely related to the pressure and material in the airbag. The polishing head strength is used to characterize the surface elastic properties of the polishing head. The polishing head rotation speed refers to the rotation speed of the polishing head during the polishing process. Since the roughness cannot be detected in real time during the polishing process, the rotation speed remains constant after setting. The curvature reflects the geometric shape of the polishing surface and is an important indicator for measuring the contact condition between the polishing head and the workpiece during polishing. The polishing liquid abrasive concentration is related to the properties of the polishing liquid. The abrasive is a particle used to cut the workpiece in the polishing liquid. The higher the concentration, the stronger the polishing ability, making it easier to achieve the desired polishing effect.
[0113] The classification method of polishing grade can be flexibly set according to actual needs, but must strictly follow the standard of smoothness. In this scheme, the polishing grade is measured and set according to the surface roughness level value. The measurement method of polishing grade also belongs to the prior art and will not be repeated here. In practice, the polishing grade can be reflected according to the reflection information of light on the surface or the transmission information on transparent objects.
[0114] In addition, this solution does not take polishing time into consideration because the contact time between the polishing head and each polishing area during the polishing process is basically the same, and each area is polished only once. After each polishing is completed, the polishing level will be measured and recorded. For example, for a certain polishing area, after the polishing data is set, the polishing head will traverse the entire area at one time for polishing. After the polishing is completed, the polishing level is checked. If it meets the requirements, the polishing is stopped, otherwise the polishing data is reset and polishing is performed again. This solution has introduced the model structure of the roughness prediction model, the meaning of each element in the polishing data, and the setting method of the polishing level. Given that the polishing level is affected by the polishing data, rather than the polishing level affecting the polishing data, the roughness prediction model can use the polishing data as input to predict the polishing level. This solution also needs to be able to infer the polishing data based on the polishing level. To this end, this application proposes the following technical solutions:
[0115] The prediction device includes an information integration module;
[0116] The information integration module is connected to the prediction module signal, and is used to obtain the corresponding relationship between the polishing data and the roughness level in the prediction module, and screen out the predicted polishing data based on the input roughness level.
[0117] Specifically, the information integration module is essentially a reverse solution module, and its function is to select the most suitable polishing data from the various polishing data combinations that can be generated by the roughness prediction model during operation. For example, when polishing data with a polishing level of 10 is required, the information integration module will select a series of polishing data that may lead to a polishing level of 10, and input these data into the roughness prediction model. Subsequently, based on whether the polishing level predicted by the roughness prediction model is 10, it is determined whether the set of polishing data can be used as the expected polishing data.
[0118] Furthermore, the information integration module selects the expected polishing data based on the following steps:
[0119] S01: Generate a polishing information pre-selection matrix K based on the input target polishing level;
[0120]
[0121] The rows in the polishing information pre-selection matrix K represent the values of each polishing data, and the columns in the polishing information pre-selection matrix K represent the types of polishing data; n represents the total number of types of polishing data, k 1,1 Indicates the first value in the first type of polishing data; k 1,… Indicates the last value in the first type of polishing data; k n,… Indicates the last value in the nth type of polishing data.
[0122] When constructing the polishing information preselection matrix K, it can be set according to the previously predicted relatively close polishing level. For example, if you need to obtain polishing data of polishing level 9, and polishing data of polishing levels 8 and 10 have been predicted before, the minimum value of these two polishing data can be used as the lower limit and the maximum value as the upper limit to generate the preselection matrix. Specifically, if the polishing head speed corresponding to polishing level 8 is 500rpm / min, and the polishing head speed corresponding to polishing level 9 is 700rpm / min, then in the polishing information preselection matrix K, the value range of the polishing head speed is set to between 500rpm / min and 700rpm / min. As for the spacing between two adjacent elements, it can be determined according to pre-set rules. Experience shows that if the polishing head speed is to have a significant impact on the polishing level, it must be increased by at least 100rpm / min, so a value less than 100rpm / min can be set as the spacing between two adjacent elements.
[0123] The above is the specific method of generating the polishing information pre-selection matrix. It can be foreseen that by selecting an element in each row of the polishing information pre-selection matrix K and inputting the polishing data composed of these elements into the roughness prediction model, a polishing level can be predicted. If the polishing level is equal to or close to the target polishing level, it means that the set of polishing data is valid. As for how to select the best element combination, this scheme is implemented by configuring the objective function f(x) and the particle swarm algorithm iteration.
[0124] The details are as follows:
[0125] S02: configure the objective function f(x);
[0126] f(x) = Q1 + Q2;
[0127] Among them, x is the input of the objective function f(x), Q1 is the polishing data change score, which is negatively correlated with the polishing data change rate; Q2 is the polishing data close score;
[0128]
[0129] Among them, w n Represents the weight score of the nth polishing data, r n Indicates the change rate of the nth polishing data under the current value;
[0130]
[0131] Among them, a represents the number of iterations index, Y a It represents the polishing level predicted by the selected polishing data input into the roughness prediction model at the ath iteration, and Y0 represents the target polishing level.
[0132] f(x) in this scheme is related to Q1 and Q2. Q1 is used to describe the degree of change in the polishing data, that is, when updating the feasible solution, the changes to the polishing data are minimized. The reason for this is to reduce the waiting time for polishing work. In practice, the abrasive concentration in the polishing liquid has little effect on the polishing level, and it is more of an overlapping effect. When the abrasive concentration in the polishing liquid changes, it is necessary to configure the polishing liquid or re-prepare the polishing liquid, which takes a long time. Therefore, the weight score of the abrasive concentration of the polishing liquid is greater than the weight score of the remaining polishing data. In practice, if the polishing data change rate is large (the polishing data changes dramatically), the larger the denominator in Q1, the smaller Q1. Q2 is the degree of closeness between the polishing data and the target polishing data. If the two are closer, Q2 is larger. Therefore, the objective function set in this scheme can update the polishing data as much as possible in the direction of approaching the target polishing level and reducing the change of the polishing data during iteration.
[0133] S03: Preselect matrix K based on polishing information to construct feasible solution matrix D, set the maximum number of iterations, number of particles, particle initial velocity position x ij ;
[0134]
[0135] Among them, x ij is the value of the jth dimension of the ith particle, i∈{1,2,…C}, j∈{1,2,…h}, Y(0,1) is used to generate any random number between 0 and 1, h represents the dimension of the feasible solution matrix D, and Respectively represent x ij The maximum and minimum values of ;
[0136] The feasible solution matrix D is constructed by selecting an element in each row of the information pre-selection matrix K. The particles in this scheme are randomly diffused in the feasible solution matrix D. The positions of the particles will diffuse in each iteration. The particles are affected by the objective function during diffusion, so as to find the best particle position. Each particle position indicates a polishing data value in the information pre-selection matrix K.
[0137] S04: Update the particle position based on the maximum objective function, repeat the above steps, and perform m iterations continuously until the maximum number of iterations is reached or Q2 reaches the threshold.
[0138] The above scheme is based on the particle swarm algorithm to select an element in each row of the polishing information pre-selection matrix K to generate polishing data. Because the particle swarm algorithm considers the need to make the predicted polishing level close to the target polishing level during iteration, the particle swarm will try to filter out polishing data in this direction when updating and iterating, and then find suitable polishing data with as few iterations as possible (fewer calculations of the roughness prediction model).
[0139] Step 5: Input the polishing plan into the roughness prediction model to generate predicted polishing data, and control the polishing operation based on the predicted polishing data.
[0140] Step 3 mainly explains the information that needs to be collected during polishing and the update and training of the roughness prediction model. Step 4 mainly explains how to build and train the roughness prediction model. Step 5 is based on steps 3 and 4. When the roughness prediction model has been built, the polishing area is polished in turn according to the polishing plan. Since the polishing data is controlled by the pre-trained roughness prediction model, the polishing level of the surface during workpiece polishing can be effectively controlled to increase the processing accuracy.
[0141] refer to Figure 2 , Example 2:
[0142] In the polishing operation control method provided in Example 1, the key lies in the prediction accuracy of the roughness prediction model, and the prediction accuracy of the roughness prediction model is affected by the network structure. Generally speaking, the more complex the network structure is, the more prediction data is, and the higher the prediction accuracy is. However, the more complex the network structure is, the more data is input, and the longer the required calculation time is. For this reason, Example 2 provides a training method for a roughness prediction model based on Example 1.
[0143] Specifically, the following steps are performed when training the roughness prediction model:
[0144] S1: construct a teacher model and train the teacher model with complete modeling data, wherein the complete modeling data includes polishing data, polishing trajectory data, free abrasive particle impact data, effective abrasive particle impact data, and abrasive particle and workpiece elasticity data;
[0145] S2: Input the training samples into the teacher model and the roughness prediction model respectively, use the loss between the roughness prediction model and the teacher model as the distillation loss, and update the built-in parameters of the roughness prediction model;
[0146] Among them, the roughness in the training sample is the prediction information, and the polishing data is the input information.
[0147] In this embodiment, the network type of the teacher model and the roughness prediction model is the same, both of which belong to convolutional network models. For the input polishing data, it needs to be converted into a matrix representation. The conversion method is the existing technology and will not be repeated here.
[0148] For the teacher model, it is trained on the basis of polishing data using more complex polishing trajectory data, free abrasive impact data, effective abrasive impact data, and abrasive and workpiece elasticity data.
[0149] Specifically, the polishing trajectory data represents the coordinate information of the path of the polishing head during the polishing process. During the polishing operation, the polishing head usually works along a straight trajectory, but due to factors such as vibration, the actual motion path will present a curved feature and affect the polishing effect. The acquisition of the polishing trajectory depends on the displacement sensor and gyroscope equipped by the polishing head.
[0150] The reason why this scheme needs to obtain the polishing trajectory is that there are differences in the cutting effect of the polishing head on the workpiece surface when it moves in a straight line and in a curve. The main reason is that in a straight path, the polishing head mainly contacts the workpiece with the front face, while in a curved path, the polishing head contacts the workpiece with the arc face. Although the polishing path is the same, the number of times the polishing head contacts the workpiece is different, and due to the existence of the inclination angle, the depth of abrasive cutting will also be different.
[0151] The free abrasive particle effect data is mainly used to describe the cutting effect of the free abrasive particles in the polishing liquid on the workpiece. The traditional view is that the cutting effect can only be achieved when the abrasive particles act between the polishing head and the workpiece. In fact, the free abrasive particles in the polishing liquid will also have a slight cutting effect on the workpiece surface, thereby affecting its flatness. To obtain the free abrasive particle effect data, it is only necessary to measure the surface density of the polishing abrasive particles in the polishing liquid and the liquid flow rate of the polishing liquid.
[0152] The effective abrasive impact data is the opposite of the free abrasive impact data. It is used to describe the cutting effect of the polishing head on the workpiece when the abrasive is squeezed. The effective abrasive impact data is obtained by calculating the product of the surface density of the polishing abrasive in the polishing liquid and the contact area between the polishing head and the workpiece.
[0153] The elasticity data of abrasive and workpiece are mainly used to describe the cutting ability of abrasive on workpiece, including the radius of abrasive, elastic coefficient of workpiece, elastic coefficient of polishing head and elastic coefficient of abrasive. This information belongs to fixed value or measurement value. If the radius of abrasive is not changed, the type of abrasive will not change. The elastic coefficient of abrasive and the elastic coefficient of workpiece need to be measured in advance. The elastic coefficient of polishing head is related to the airbag pressure and the material of polishing head, which also needs to be measured in advance, and the elastic coefficient of polishing head under different pressures should be recorded.
[0154] The above are the sources or acquisition methods of all data in the complete modeling data. In essence, the complete modeling data adds more detailed information on the basis of polishing data. Among them, polishing data is data that is easy to change during the polishing process, or it is empirical data, which is most likely to affect the polishing grade after polishing. In the complete modeling data, in addition to polishing data, other data are some data with more complex influences or difficult to change. The complete modeling data is actually the indispensable information when establishing a polishing and cutting model under microscopic conditions. Therefore, using complete modeling data to train the teacher model can make the teacher model have higher prediction accuracy.
[0155] In the technical solution provided by the present application, a teacher model is pre-constructed when training the roughness model. The teacher model is trained using more complete and complex full modeling data, so that its prediction accuracy is higher. Then, the teacher model is used to perform distillation learning on the roughness prediction model. In this way, the roughness prediction model can learn the teacher model's ability to understand the training samples even when the model complexity is low. Therefore, even using empirical data (polishing data), a good prediction accuracy can be achieved.
[0156] The roughness prediction model needs to predict the polishing level after polishing based on the polishing data. In the polishing data, the curvature of the polishing area is difficult to change. When using the roughness prediction model, it is a fixed value, while other polishing data are values that need to be adjusted. For this reason, when training the roughness prediction model, it is necessary to explore the influence of the curvature of the polishing area on the polishing level. To this end, the present application provides the following technical solutions: the roughness prediction model includes a first student network, a second student network and a fusion output layer, and the first student network and the second student network are respectively connected to the fusion output layer signal. The first student network inputs the polishing head strength, the polishing head rotation speed, and the polishing liquid abrasive concentration, and the second student network inputs the curvature of the polishing area. The output of the first student network and the output of the second student network are used as the input of the fusion output layer, and the fusion output layer finally outputs the polishing level.
[0157] In this scheme, the roughness prediction model is divided into two independent networks, and the final output results of the two networks are fused to output the polishing grade. During the training process, the two student networks can learn the correspondence between polishing data and polishing grades in the teacher model.
[0158] In this scheme, both the first student network and the second student network are trained by the teacher model at the same time.
[0159] The teacher model includes:
[0160] Input layer: input complete modeling data;
[0161] Convolutional layer: multiple layers are set to capture the features of input data;
[0162] Pooling layer: used to reduce the size of features, reduce the amount of calculation, and retain important features;
[0163] Fully connected layer: used to integrate global features;
[0164] Output layer: output classification results (polishing level);
[0165] The structures of the first student network and the second student network are the same. Therefore, only the structure of the first student network is explained here.
[0166] Specifically, the First Student Network includes:
[0167] Input layer: Same as the teacher model. Here only the polished data needs to be input.
[0168] Convolutional layers: The number is smaller than that of the teacher model.
[0169] Fully connected layer: used to integrate global features.
[0170] Output layer: output hidden features;
[0171] The following is the training process:
[0172] S1: Select the convolutional layers related to the first student network and the second student network from the teacher network as the distillation training layers.
[0173] Specifically, the first student network selects the high-level convolutional layer in the teacher model as the distillation training layer, and the second student network selects the low-level convolutional layer in the teacher model as the distillation training layer. The teacher model in this scheme has 100 hidden layers, the low-level convolutional layer is the 10th layer, and the high-level convolutional layer is the 90th layer.
[0174] The reason for this setting is that the second student network is mainly responsible for extracting the influence of the curvature of the polishing area and the polishing level. This feature is relatively simple and the influence is not complicated. The reason is that the curvature of the polishing area will directly affect the contact area between the polishing head and the workpiece. If the contact area is different, the micro-cutting model during polishing will change directly, so the curvature of the polishing area is actually the low-level classification information. Different curvatures require the re-establishment of the cutting model, so it is a low-level feature in the neural network model and is mainly processed in the low-level convolutional layer of the teacher model. On the contrary, the first student network inputs the remaining polishing data, which is a high-level feature and needs to correspond to the high-level convolutional layer in the teacher model.
[0175] S2: Define the loss function L;
[0176] L = L1 + L2 + L3;
[0177] L3 is the fusion output layer loss, which is used to describe the loss between the final fusion output layer output result and the actual result. L2 is the distillation loss between the second learning network and the distillation training layer, and L1 is the distillation loss between the second learning network and the distillation training layer. The above three loss functions can choose the cross entropy loss function or the mean square error (MSE) loss function.
[0178] S3: Keep the weights of the teacher model unchanged and only update the weights of the student network; through multiple iterative training, the student network gradually learns the distilled training layer representation of the teacher network while retaining the ability to learn classification tasks.
[0179] Furthermore, during training, the teacher model needs to input complete modeling data based on the student network, that is, the teacher model inputs complete modeling data and the student network inputs polished data.
[0180] Embodiment 3:
[0181] A polishing operation control system is provided, comprising:
[0182] An information input module, used to input an expected polishing grade and polishing plan;
[0183] A prediction device, wherein the prediction device has a built-in roughness prediction model;
[0184] A control module controls the polishing level based on the polishing operation control method.
[0185] The above description is only some preferred embodiments of the present application and an explanation of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present application is not limited to the technical solutions formed by the specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above invention concept. For example, the above features are replaced with (but not limited to) the technical features with similar functions disclosed in the embodiments of the present application to form a technical solution.
Claims
1. A polishing operation control method, characterized in that: The steps include: Step 1: Obtain three-dimensional information of the workpiece surface and generate a polishing area on the workpiece surface; Step 2: Obtain polishing requirements for each polishing area, and generate a polishing plan based on the polishing requirements; Step 3: Polish the workpiece based on the polishing plan and collect polishing data in real time; The polishing data includes the strength of the polishing head, the rotation speed of the polishing head, the curvature of the polishing area and the abrasive concentration of the polishing liquid; Step 4: Setting a prediction device, which has a built-in roughness prediction model; Step 5: Input the polishing plan into the roughness prediction model to generate predicted polishing data, and control the polishing operation based on the predicted polishing data.
2. The polishing operation control method according to claim 1, characterized in that: Step 1 includes the following steps: Step 11: Identify all areas on the workpiece that need to be polished and use them as polishing areas. Step 12: Collect three-dimensional information of the surface of the polishing area and calculate the curvature of the polishing area.
3. The polishing operation control method according to claim 1, characterized in that: Step 2 includes the following steps: Step 21: Calculate the difference between the initial polishing level and the target polishing level of each polishing area to obtain a mark value of each polishing area; Step 22: sort all polishing areas according to the mark value from small to large to form a polishing area sequence; Step 23: Polish the polishing area with a small mark value first, and then polish the polishing area with a large mark value to generate a polishing plan.
4. The polishing operation control method according to claim 1, characterized in that: Step 3 includes the following steps: Step 31: Based on the polishing plan, determine the execution order of each polishing area on the workpiece, and polish each polishing area according to the execution order; Step 32: After polishing of each polishing area is completed, the polishing level of the polishing area and the polishing data of this polishing are obtained, and the obtained polishing level and polishing data of this polishing are input into the prediction device.
5. The polishing operation control method according to claim 4, characterized in that: The prediction device includes: Prediction module, with built-in roughness prediction model for predicting roughness; The updating module receives the roughness level of the polishing area after polishing and the polishing data of this polishing, and inputs them into the prediction module to feedback and update the roughness prediction model.
6. The polishing operation control method according to claim 1, characterized in that: The following steps are performed when training the roughness prediction model: S1: construct a teacher model and train the teacher model with complete modeling data, wherein the complete modeling data includes polishing data, polishing trajectory data, free abrasive particle impact data, effective abrasive particle impact data, and abrasive particle and workpiece elasticity data; S2: Input the training samples into the teacher model and the roughness prediction model respectively, use the loss between the roughness prediction model and the teacher model as the distillation loss, and update the built-in parameters of the roughness prediction model; Among them, the roughness in the training sample is the prediction information, and the polishing data is the input information.
7. The polishing operation control method according to claim 5, characterized in that: The prediction device includes an information integration module; The information integration module is connected to the prediction module signal, and is used to obtain the corresponding relationship between the polishing data and the roughness level in the prediction module, and screen out the predicted polishing data based on the input roughness level.
8. The polishing operation control method according to claim 7, characterized in that: The information integration module selects the expected polishing data based on the following steps: S01: Generate a polishing information pre-selection matrix based on the input target polishing level; The rows in the polishing information pre-selection matrix K represent the values of each polishing data, and the columns in the polishing information pre-selection matrix K represent the types of polishing data; n represents the total number of types of polishing data, k 1,1 Indicates the first value in the first type of polishing data; k n,… Indicates the last value in the nth polishing data; S02: configure the objective function f(x); f(x)=Q1+Q2; Among them, x is the input of the objective function f(x), Q1 is the polishing data change score, which is negatively correlated with the polishing data change rate; Q2 is the polishing data close score; Among them, w n Represents the weight score of the nth polishing data, r n Indicates the change rate of the nth polishing data under the current value; Among them, a represents the number of iterations index, Y a represents the polishing level predicted by the selected polishing data input into the roughness prediction model at the ath iteration, and Y0 represents the target polishing level; S03: Preselect matrix K based on polishing information to construct feasible solution matrix D, set the maximum number of iterations, number of particles, particle initial velocity position x ij ; Among them, x ij is the value of the jth dimension of the ith particle, and Respectively represent x ij The maximum and minimum values of , i∈{1,2,…C}, j∈{1,2,…h}, Y(0,1) is used to generate any random number between 0 and 1, and h represents the dimension of the feasible solution matrix D; S04: Update the particle position based on the maximum objective function, repeat the above steps, and perform m iterations continuously until the maximum number of iterations is reached or Q2 reaches the threshold.
9. The polishing operation control method according to claim 8, characterized in that: The weight score of polishing liquid abrasive concentration is greater than the weight scores of other polishing data.
10. A polishing operation control system, characterized in that: include: An information input module, used to input an expected polishing grade and polishing plan; A prediction device, wherein the prediction device has a built-in roughness prediction model; A control module controls the polishing level based on the polishing operation control method according to any one of claims 1 to 9.