XY motion platform positioning error compensation method and system

Abstract

The invention relates to the technical field of precise motion control, in particular to an XY motion platform positioning error compensation system and method, which extracts spatial-temporal characteristics through multi-modal data fusion and normalization processing in combination with a neural network hybrid model, and dynamically manages time sequence errors by using a forgetting gate, an input gate and an output gate. And a compensation parameter is updated by adopting an online adaptive training mechanism of error source classification. The problems that in the prior art, due to mechanical abrasion and thermal deformation of an encoder, precision is attenuated, nonlinear errors are difficult to process through a PID algorithm, pure vision positioning is prone to interference and complex in calibration, and an online learning mechanism is lacked can be effectively solved, the positioning comprehensive error is reduced to the micron order, the anti-interference robustness and the real-time compensation capacity of a system are improved, and the system reliability is improved. And the positioning precision and the production efficiency are obviously improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of precision motion control, and particularly relates to an XY motion platform positioning error compensation method and system. BACKGROUND

[0002] In the semiconductor industry, especially in the chip packaging manufacturing link, the positioning motion precision and efficiency in the packaging process are crucial to production. The positioning precision directly affects the performance and reliability of the chip, and the positioning efficiency determines the packaging production efficiency. Therefore, realizing the precise control and error compensation of the positioning motion of the XY platform of the packaging equipment is a key link in the chip packaging process. However, the positioning scheme of the XY platform in the prior art has obvious limitations: 1. The feedback mode relying on the encoder is easily affected by mechanical wear, thermal deformation and other factors, resulting in long-term precision decline; 2. The general control algorithm based on PID or feedforward compensation is difficult to effectively handle nonlinear errors such as friction, inertia and hysteresis; 3. The pure visual positioning scheme is not only easily disturbed by environmental factors such as light and shielding, but also the calibration process is complex and cannot avoid the influence of the inherent nonlinear error of the structure.

[0003] Therefore, the present application is proposed. SUMMARY

[0004] In order to solve the above technical problems in the prior art, the present application provides an XY motion platform positioning error compensation method and system, which can solve the problems that the feedback is easily affected by mechanical wear, thermal deformation, and it is difficult to handle nonlinear errors such as friction and inertia hysteresis in the prior art.

[0005] To achieve the above purpose, the technical scheme of the present application is as follows: In a first aspect, an XY motion platform positioning error compensation method, the specific steps comprising: Synchronously collecting multi-modal data after motion positioning, differentiating and normalizing the multi-modal data, and constructing a space-time feature vector; extracting spatial correlation features and time sequence evolution features through a network hybrid model, and calculating an output error compensation amount.

[0006] Further, the multi-modal data includes: X-axis image position deviation, Y-axis image position deviation, X-axis encoder position deviation, Y-axis encoder position deviation, displacement of positioning motion, velocity of positioning motion, acceleration of positioning motion and motor current of positioning motion.

[0007] Further, the X-axis image position deviation, Y-axis image position deviation, X-axis encoder position deviation and Y-axis encoder position deviation are normalized, and the specific formula is:

[0008] wherein, is the original bias data, is the minimum range boundary of the original data, is the maximum range boundary of the original data.

[0009] Further, the displacement of the positioning motion is logarithmized and then normalized, and the logarithm calculation formula is:

[0010] wherein, is the original displacement; is the result of logarithm calculation; and the calculation formula of the normalization processing is:

[0011] wherein, and are the minimum and maximum boundary values of L respectively; is the new displacement value obtained after the normalization processing of the formula; The motor current of the positioning motion is logarithmized and then normalized, and the logarithm calculation formula is:

[0012] wherein, is the original displacement; is the result of logarithm calculation; and the calculation formula of the normalization processing is:

[0013] wherein, is the minimum boundary value; is the maximum boundary value; and the taking is ; The speed of the positioning motion and the acceleration of the positioning motion are standardized, and the standardization formula is as follows:

[0014] wherein, is the mean value, is the standard deviation.

[0015] Further, the network hybrid model is constructed by using a convolutional neural network and a long short-term memory network; The spatial correlation feature is extracted along the parameter dimension by the convolutional neural network branch, and the convolution operation is: ​​

[0016] wherein, is a typical nonlinear activation function; is the convolution kernel weight corresponding to the input channel; is the number of input feature channels; is the feature vector of the input channel at time t; is the bias term; The time evolution feature is extracted along the time dimension by a long short-term memory network branch, and the specific calculation process is as follows:

[0017]

[0018] wherein, is an element-wise multiplication; is a candidate memory cell in the long short-term memory network; is a memory cell in the long short-term memory network; is a forget gate in the long short-term memory network; is an input gate in the long short-term memory network; is an output gate in the long short-term memory network; is a hidden state in the long short-term memory network.

[0019] Further, the specific calculation formulas of the forget gate, the input gate and the output gate are as follows:

[0020]

[0021]

[0022]

[0023] wherein, and tanh are core activation functions in deep learning; , , , is a weight matrix to be learned; , , , is a bias vector to be learned; is a vector obtained by splicing the hidden state at time t-1 and the convolution output .

[0024] Further, the error compensation amount is calculated by the following formula:

[0025]

[0026] wherein, is the hidden state of the long short-term memory network is the feature sub-vector related to the X-axis extracted from the hidden state of the long short-term memory network; is the feature sub-vector related to the Y-axis extracted from the hidden state of the long short-term memory network; is the training weight vector of the X-axis component; is the training weight vector of the Y-axis component; is the training scalar bias term of the X-axis component; is the training scalar bias term of the Y-axis component.

[0027] Further, the training weight vector and the training scalar bias term are updated adaptively, and the specific formula is as follows:

[0028] wherein, is the X-axis weight vector is the value of the i-th element of the X-axis weight vector at the t-th iteration; is the weight learning rate; is the bias learning rate; is the current X-axis residual error; is the X-axis bias term is the value of the X-axis bias term at the t-th iteration; is the mathematical sign function.

[0029] In a second aspect, an XY motion platform positioning error compensation system includes: an upper computer, a motion controller, an IO execution element, X\Y\Z-axis servo drivers, X\Y\Z-axis linear motors, X\Y\Z-axis grating rulers, and an industrial camera. The upper computer is connected with the motion controller and the industrial camera through network cables respectively. The motion controller is connected with the X\Y\Z-axis servo drivers through network cables respectively. The motion controller is connected with the X\Y\Z-axis grating rulers through level signal cables respectively. The X\Y\Z-axis servo drivers are connected with the corresponding linear motors through analog voltage signal cables respectively.

[0030] Further, the upper computer includes: a parameter management module for setting, saving, and managing system parameters; ​A file management module for processing file, parameter file and state log file management; A process action module for sending instructions to the motion controller according to the processing process, controlling the linear motor motion and IO execution element action; A visual processing module for image acquisition, management and visual deviation calculation; A positioning compensation module for running an error compensation method, outputting a positioning error compensation amount to the process action module for displacement compensation when positioning motion is performed; A display management module for interface display and interaction management of the software.

[0031] Compared with the prior art, the XY motion platform positioning error compensation method and system provided by the application can improve the positioning accuracy and positioning efficiency of the equipment packaging motion, thereby improving the overall production quality and production efficiency, reducing the production cost, and enhancing the market competitiveness of the product. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1 The flowchart of the XY motion platform positioning error compensation method provided by the first embodiment of the application; Figure 2 The architecture diagram of the XY motion platform positioning error compensation system provided by the first embodiment of the application. DETAILED DESCRIPTION

[0033] The technical solutions of the application will be described clearly in the following description with reference to the accompanying drawings. Obviously, the described embodiments are not all the embodiments of the application, and all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the protection scope of the application.

[0034] It should be noted that, unless otherwise specified, the relative arrangement of the components and steps, numerical expressions described in these embodiments should not be understood as limiting the scope of the application.

[0035] The following description of exemplary embodiments is merely illustrative in nature and is in no way intended to limit the application or its application or use in any way. Techniques, methods and devices known to those of ordinary skill in the relevant art may not be discussed in detail here, but in the case of applicable techniques, methods and devices, these techniques, methods and devices should be considered as part of this specification.

[0036] Embodiment one Reference Figure 1 , Figure 1 A flow chart of an XY motion platform positioning error compensation method is provided for the present application, and the specific steps include: S1, after motion positioning, synchronously collect multi-modal data; after the XY platform completes positioning, obtain the current position image of the platform through the machine vision module, extract the actual coordinates and compare them with the target coordinates to generate an image position deviation and , obtain the deviation between the encoder feedback position and the target position through the motion controller and , obtain the displacement of this positioning motion through the motion control module , speed , acceleration and motor current ; Among them, the image position deviation reflects the visual detection error of platform positioning, which is used to evaluate the deviation degree of actual position and theoretical target; the encoder position deviation reflects the position error of the motor driving system, which is the direct data of the motion control closed loop feedback; the displacement measures the size of the motion stroke, which affects the positioning time and energy consumption and other parameters; speed and acceleration reflect the dynamic characteristics of motion, speed affects the positioning efficiency, and acceleration affects the motion stability and impact; the motor current reflects the size of the motor load, and the larger the current usually indicates the larger the load or resistance, which can be used for fault diagnosis or energy consumption analysis.

[0037] S2, normalize the above parameters, which can specifically include: S21, normalize the image position deviation , and the encoder position deviation and The calculation formula is:

[0038] Among them, is the original deviation data, is the minimum value range boundary of the original data, is the maximum value range boundary of the original data. Take , as , normalize the image position deviation , and the encoder position deviation and to [0, 1].

[0039] S22, for displacement​ Take logarithm Normalization processing, taking logarithm can effectively process large range data fluctuations, and the logarithm calculation formula is:

[0040] Among them, is the original displacement; Take logarithm calculation results. Then Normalization, the calculation formula is:

[0041] Among them, and are the minimum and maximum boundary values of L set respectively; is the new displacement value obtained after normalization processing by the formula; take ] for Similarly, the current I also takes logarithm after Normalization processing, and the logarithm calculation formula is:

[0042] Among them, is the original displacement; is the logarithm calculation result; then Min-Max normalization, the calculation formula is:

[0043] Among them, is the minimum boundary value; is the maximum boundary value; take ] for .

[0044] S23, the speed and acceleration Take Standardization, The standardization formula is as follows:

[0045] Among them, is the mean, is the standard deviation, take the mean and standard deviation of the speed , , the mean and standard deviation of the acceleration , .

[0046] S3, construct the space-time feature vector of the above normalized parameters; S31, according to the above parameters, take the feature vector of each time step as: ; S32. Define the time window length, i.e., the number of time steps, as 4, and perform spatiotemporal feature concatenation. After flattening, the feature dimension is 4×8=32, as follows:

[0047] S4. Construct a convolutional neural network-long short-term memory network (CNN-LSTM) hybrid model and calculate the output positioning error compensation.

[0048] S41, the convolutional neural network (CNN) branch extracts spatial correlation features along the parameter feature dimension. It contains a one-dimensional convolution layer, which takes the above feature vector as input. The convolution operation is as follows:

[0049] in, Typical Non-linear activation function; For the The convolution kernel weights corresponding to the input channels; is the number of channels of the input feature; For the The feature vector of the input channel at time t; is the bias term; Similarly, the weight is not a pre-set fixed value, but a parameter automatically learned through the training process. It is a trainable parameter matrix and is part of the model.

[0050] During the optimization process, the image position deviation is automatically updated by gradient descent. 、 and encoder position deviation 、 , take the initial value , for displacement ,speed and acceleration , take the initial value , for the motor current , take the initial value ; is the eigenvector of the i-th sensor at time step t; b is the bias term, which compensates for the inherent deviation of the system and has a value of 0.01 to enhance the robustness of the model to the zero drift of the system; is the number of input channels, which is 8 in this system; is the convolution output feature map at time step t, Dimension, The initial value of the number of output channels is 32, and T is the time step length, which is 4ms.

[0051] S42, a long short-term memory (LSTM) branch extracts time evolution features along the time dimension, and a convolutional neural network (CNN) extracts spatial features with the historical hidden state input LSTM unit to generate dynamic error coding The core calculation process of LSTM is as follows:

[0052]

[0053] wherein, is an element-wise multiplication; is a candidate memory cell in LSTM; is a memory cell in LSTM; is a forget gate in LSTM; is an input gate in LSTM; is an output gate in LSTM; is a hidden state in LSTM.

[0054] Specifically, the candidate memory cell is generated from the current input, contains new information at the current time step, and is activated by a tanh function, the value of which is between -1 and 1, which represents the new information extracted from the current sensor feature and the previous hidden state.

[0055] The memory cell is responsible for transmitting information in the time series, which embodies the long-term state information of the system such as position error accumulation, vibration and mechanical deformation, etc.

[0056] The forget gate controls how much information in the previous memory cell state is retained in the current memory cell , which determines which information in the past is related to the current state.

[0057] The input gate controls how much information in the current candidate memory needs to be updated to the memory cell, which determines the degree of contribution of the current input to the update of the memory cell.

[0058] The output gate controls how much information in the memory cell is output to the hidden state , which determines which information in the current memory cell is useful for the prediction (or compensation) of the next time step.

[0059] The hidden state is the output of LSTM to the outside, which contains the information that needs to be transmitted to the next time step at the current time step and the output at the current time step, which is calculated based on the memory cell and the output gate, and encodes the filtered state information (such as compensation control quantity) useful for compensation control. The dimension is empirically taken as 96.

[0060] The specific calculation formulas of the forget gate, the input gate, the output gate and the cell in the above LSTM are as follows:

[0061]

[0062]

[0063]

[0064] wherein, and tanh are core activation functions in deep learning; , , , is a weight matrix to be learned; , , , is a bias vector to be learned; is a vector obtained by splicing the hidden state at time t-1 and the convolution output at time t; all the weight matrices and the bias vector are randomly initialized, and effective weight and bias data are obtained by self-learning adjustment during training and use.

[0065] is a function for compressing the output to the interval (0, 1), and the mathematical expression is as follows:

[0066] wherein, is an operation item of the natural exponential function. tanh is a hyperbolic tangent function, and the output range is (-1, 1), and the mathematical expression is as follows:

[0067] wherein, and are operation items of the natural exponential function.

[0068] S43, according to the hidden state obtained above, the respective positioning compensations of the XY axes are obtained according to the following calculation: and The specific calculation formula is as follows:

[0069]

[0070] wherein, is the LSTM hidden state is the feature sub-vector related to X-axis (first 48 dimensions) extracted from the LSTM hidden state = [0:47]; is the feature sub-vector related to Y-axis (last 48 dimensions) extracted from the LSTM hidden state is the feature sub-vector related to Y-axis (last 48 dimensions) extracted from the LSTM hidden state = [48:95]; is the training weight vector of X-axis component; is the training weight vector of Y-axis component; is the training scalar bias term of X-axis component; is the training scalar bias term of Y-axis component; and are obtained by training, and taking X-axis training as an example, specifically comprising: S431, initializing from physical meaning into three error sources, and setting , k: error source three types (structural deformation / vibration / inertia), then is the corresponding hidden state component, and k divides the dimension of into three parts; According to experience, the initial values are: structural deformation error source component , vibration error source component and inertia error source component ; The initial X-axis zero bias is obtained by using a laser interferometer.

[0071] S432, adding training constraints to , according to three error sources, and according to the empirical values as follows: , , .

[0072] S433, training according to the following online adaptive mechanism calculation method, and determining the weight and the bias term according to the best ; specifically comprising:

[0073] wherein, is the value of the th element of the X-axis weight vector at the tth iteration; is the weight learning rate; is the bias learning rate; is the current X-axis residual error; is the X-axis bias term is the value at the t-th iteration; is the mathematical sign function when > 0, , , , < 0, .

[0074] In a second aspect, referring to Figure 2 The XY motion platform positioning error compensation system provided by the application comprises a host computer, a motion controller, IO execution elements, X / Y / Z axis servo drivers, X / Y / Z axis linear motors, X / Y / Z axis grating rulers and an industrial camera, wherein the host computer adopts a PC computer and runs a host computer control program, is connected with the motion controller and the industrial camera through network cables respectively, the motion controller is connected with the X / Y / Z axis servo drivers through network cables based on an EtherCat bus protocol and is connected with the X / Y / Z axis grating rulers based on TTL level signal cables, and the X / Y / Z axis servo drivers are connected with the corresponding axis linear motors through analog voltage signal cables.

[0075] The host computer control program comprises six software function modules, and specifically can comprise: A parameter management module is used for realizing the setting, saving and management of all parameters of the system through a graphical interface. A file management module is used for managing processing files, parameter files and state log files in a database mode. A process action module is used for generating a motion trajectory according to a processing process and sending an instruction to the motion controller to control the linear motor and the IO execution elements to act. A display management module is used for realizing the display and interactive management of the software interface by using a multi-window layout. A positioning compensation module is used for calling an error compensation method to calculate a positioning error compensation amount and outputting the positioning error compensation amount to the process action module. A visual processing module is used for controlling the industrial camera to collect images, processing the images and calculating a visual deviation. When the system is running, the positioning compensation module realizes the XY motion platform positioning error compensation method, outputs the calculated positioning error compensation amount to the process action module, and the process action module superimposes the compensation amount when planning a positioning motion trajectory, so that the real-time compensation of the positioning error is realized.

[0076] In summary, the application has the following advantages: 1. The multi-modal data fusion reduces environmental and mechanical interference and enhances the adaptability of the system. 2. Through the CNN-LSTM hybrid model and global training correction, the nonlinear error is effectively processed, the comprehensive error is reduced to micrometer level, and the high-precision requirement of chip packaging is met; 3. Through the online learning mechanism, the model parameters can be optimized in real time, the long-term drift and sudden disturbance of the equipment are coped with, and the long-term stable precision is ensured; 4. Through improving the packaging motion positioning precision and efficiency, the chip production quality, efficiency and cost are improved, and the product market competitiveness is enhanced.

[0077] The above specific embodiments are only used to illustrate the technical solutions of the present application but not limit the present application. Although the present application has been described in detail with reference to the examples, it should be understood by those skilled in the art that the technical solutions of the present application can be modified or replaced equivalently without departing from the scope of the technical solutions of the present application, and all should be covered in the scope of the claims of the present application.

Claims

1. A method for compensating positioning errors of an XY motion platform, characterized in that: The specific steps include: After motion positioning, multimodal data is synchronously collected, the multimodal data is subjected to differential normalization processing, and a spatiotemporal feature vector is constructed; spatial correlation features and temporal evolution features are extracted through a network hybrid model, and the output error compensation amount is calculated.

2. The XY motion platform positioning error compensation method according to claim 1, characterized in that: The multimodal data includes: X-axis image position deviation, Y-axis image position deviation, X-axis encoder position deviation, Y-axis encoder position deviation, positioning movement displacement, positioning movement speed, positioning movement acceleration and positioning movement motor current.

3. The XY motion platform positioning error compensation method according to claim 2, characterized in that: The X-axis image position deviation, Y-axis image position deviation, X-axis encoder position deviation, and Y-axis encoder position deviation are normalized. The specific formula is: in, is the original deviation data, is the minimum range boundary of the original data, The maximum value range boundary of the original data.

4. The XY motion platform positioning error compensation method according to claim 2, characterized in that: The displacement of the positioning motion is logarithmically normalized, and the logarithmic calculation formula is: in, is the original displacement; The result of the logarithmic calculation is performed; the calculation formula for the normalization process is: in, and are the minimum and maximum boundary values ​​of L set respectively; The new displacement value is obtained after normalization processing of the formula; The motor current of the positioning motion is taken logarithmically and then Normalized processing, the logarithmic calculation formula is: in, is the original displacement; The result of logarithmic calculation is: in, is the minimum boundary value; is the maximum boundary value; take [ ]for[ ]; The speed of the positioning movement and the acceleration of the positioning movement are adopted standardization, The normalization formula is as follows: in, is the mean, is the standard deviation.

5. The XY motion platform positioning error compensation method according to claim 1, characterized in that: The network hybrid model is constructed using a convolutional neural network and a long short-term memory network; The spatial correlation features are extracted along the parameter dimension through the convolutional neural network branch, and the convolution operation is: in, Typical Non-linear activation function; For the The convolution kernel weights corresponding to the input channels; is the number of channels of the input feature; For the The feature vector of the input channel at time t; is the bias term; The temporal evolution features are extracted along the time dimension through the long short-term memory network branch. The specific calculation process is as follows: in, is element-wise multiplication; It is a candidate memory unit in the long short-term memory network; It is the memory unit in the long short-term memory network; It is the forget gate in the long short-term memory network; It is the input gate in the long short-term memory network; It is the output gate in the long short-term memory network; is the hidden state in the long short-term memory network.

6. The XY motion platform positioning error compensation method according to claim 5, characterized in that: The specific calculation formulas for the forget gate, input gate and output gate are: in, and tanh are the core activation functions in deep learning; , , , is the weight matrix to be learned; , , , is the bias vector to be learned; To convolve the hidden state at time t−1 with the output The vector obtained by concatenation.

7. The XY motion platform positioning error compensation method according to claim 1, characterized in that: The calculation formula of the error compensation amount is: in, is the hidden state of the long short-term memory network The feature sub-vector related to the X-axis extracted from ; is the feature subvector related to the Y axis extracted from the hidden state of the long short-term memory network; is the training weight vector of the X-axis component; is the training weight vector of the Y-axis component; is the training scalar bias term of the X-axis component; is the training scalar bias term for the Y-axis component.

8. The XY motion platform positioning error compensation method according to claim 7, characterized in that: The training weight vector and the training scalar bias term are updated adaptively, and the specific formula is: in, is the X-axis weight vector No. The value of the element at the tth iteration; is the weight learning rate; is the bias learning rate; is the current X-axis residual; X-axis offset The value at the tth iteration; is a mathematical symbol function.

9. An XY motion platform positioning error compensation system, characterized in that: include: Host computer, motion controller, IO actuator, X\Y\Z axis servo drive, X\Y\Z axis linear motor, X\Y\Z axis grating ruler and industrial camera; The host computer is connected to the motion controller and the industrial camera respectively via network cables; The motion controller is connected to the X, Y and Z axis servo drivers respectively via network cables; The motion controller is connected to the X, Y and Z axis linear scales respectively based on level signal cables; The X, Y, and Z axis servo drivers are connected to the linear motors of the corresponding axes via analog voltage signal cables.

10. The XY motion platform positioning error compensation system according to claim 9, characterized in that: The host computer includes: Parameter management module, used for setting, saving and managing system parameters; File management module, used for the management of processing files, parameter files and status log files; The process action module is used to send instructions to the motion controller according to the processing technology to control the movement of the linear motor and the action of the IO actuator; Visual processing module for image acquisition, management and visual deviation calculation; A positioning compensation module, configured to execute an error compensation method and output a positioning error compensation value to the process action module for performing displacement compensation during positioning motion; Display management module, used for software interface display and interaction management.

Images