Die bonder thermal error prediction method and device, equipment and storage medium

Abstract

The invention discloses a die bonder thermal error prediction method, device and equipment and a storage medium, and relates to the technical field of data prediction, and the method comprises the following steps: collecting multiple groups of temperature data of a die bonder, and obtaining thermal errors corresponding to the multiple groups of temperature data; adding a plurality of LSTM layers and attention layers into the CNN model to obtain an initial thermal error prediction model; training the initial thermal error prediction model through the multiple groups of temperature data and the corresponding thermal errors to obtain a thermal error prediction model; inputting to-be-predicted temperature data into the thermal error prediction model, and performing feature extraction on the to-be-predicted temperature data through the CNN layer to obtain temperature features; performing time sequence feature extraction on the temperature features through a plurality of LSTM layers to obtain global features; and performing weighted summation and mapping on the global features through an attention layer to obtain a thermal error. According to the method, the CNN, the LSTM and the attention layer are combined, so that the prediction accuracy of the thermal error is greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of data prediction, and in particular to a method, device, equipment and storage medium for predicting thermal errors of a die bonder. Background Art

[0002] In recent years, my country has become the world's largest market for semiconductor equipment demand. The die bonder is one of the key core equipment in the semiconductor chip packaging and testing and light-emitting diode chip packaging and testing production lines. It is a high-speed and high-precision equipment. Improving the accuracy of the die bonder is not only an important symbol of technological innovation, but also the core driving force for promoting the high-quality development and sustainable progress of the manufacturing industry. The accuracy of the die bonder is affected by a combination of factors, including geometric error, thermal error and vibration error.

[0003] The linear die bonder is a type of die bonder. Its structure and motion are similar to those of a three-axis CNC machine or a three-coordinate measuring machine, but they differ in terms of motion characteristics and thermal characteristics. The main functions of a die bonder are to take and bond the die, emphasizing the positioning accuracy of the two positions, and having strict requirements on UPH (the number of chips produced per hour). The die bonder itself is an ultra-precision device with little influence from geometric errors. However, considering the high efficiency, high speed and rapid reciprocating motion, thermal error becomes the most important factor affecting the accuracy of the linear die bonder. A large number of studies have found that the error caused by thermal deformation accounts for 40% to 70% of the total machining error of the machine tool, and the higher the accuracy level of the machine tool, the greater the proportion of thermal error in the total error. This situation will only be more serious for linear die bonders.

[0004] In the prior art, CNN models are usually used to predict thermal errors. However, CNN memory units exhibit static characteristics, mainly focusing on feature extraction of current input data, and lack an update mechanism that changes over time. Therefore, when using CNN models to predict thermal errors, their prediction accuracy is low. Summary of the invention

[0005] The present invention provides a method, device, equipment and storage medium for predicting thermal errors of a die bonder, which solves the problem of low prediction accuracy when the existing CNN model performs thermal error prediction.

[0006] In a first aspect, the present invention provides a method for predicting thermal errors of a die bonder, comprising the following steps:

[0007] Collect multiple sets of temperature data from the die bonder and obtain thermal errors corresponding to the multiple sets of temperature data;

[0008] Add multiple LSTM layers and attention layers to the CNN model to obtain an initial thermal error prediction model; train the initial thermal error prediction model using multiple sets of temperature data and corresponding thermal errors to obtain a thermal error prediction model;

[0009] The temperature data to be predicted is input into the thermal error prediction model, and the temperature data to be predicted is subjected to feature extraction through the CNN layer to obtain temperature features. The time series features of the temperature features are extracted through multiple LSTM layers to obtain global features. The global features are weighted summed and mapped through the attention layer to obtain the thermal error.

[0010] Preferably, temperature sensors are arranged at different positions of the X-axis of the die bonding machine, each temperature sensor collects a set of temperature data, and each set of temperature data is clustered by a clustering algorithm to obtain a plurality of sensitive temperatures.

[0011] Preferably, obtaining thermal errors corresponding to multiple sets of temperature data comprises the following steps:

[0012] Selecting a starting point and an end point on the X-axis of the die bonder, and obtaining a first distance between the starting point and the end point;

[0013] Obtaining a theoretical time for the X-axis welding head to pass through the first distance according to the running speed of the X-axis welding head;

[0014] When the welding head is in operation, timing is started from the position of the starting point, and after the theoretical time has passed, the second distance between the welding head and the starting point is obtained;

[0015] When the die bonding machine is just turned on, a difference between the second distance and the first distance is obtained, where the difference is a geometric error;

[0016] Under different sensitive temperatures, the difference between the corresponding second distance and the first distance is obtained to obtain the corresponding positioning error;

[0017] Multiple positioning errors are subtracted from the geometric errors to obtain multiple thermal errors.

[0018] Preferably, before training the initial thermal error prediction model using multiple sets of temperature data and corresponding thermal errors, the method further includes optimizing the model parameters of the initial thermal error prediction model. The optimization process specifically includes the following steps:

[0019] Taking different model parameters as an individual, generating individuals with one third of the population; taking multiple individuals as the first vertex, wherein the model parameters include the number of hidden layer nodes, the initial learning rate, and the L2 regularization coefficient;

[0020] Taking the first vertices as starting points, a direction vector is defined in a spherical coordinate system, and the direction vector is subjected to coordinate transformation to obtain a second vertex; the direction vector is rotated 60 degrees counterclockwise and then subjected to coordinate transformation to obtain a third vertex; the first vertex, the second vertex and the third vertex form a triangular topological unit;

[0021] forming a fourth vertex in each triangular topological unit in a linearly weighted manner;

[0022] Obtain the fitness values ​​of all vertices in each triangular topological unit and sort them to obtain the best individual in each triangular topological unit, and randomly select the best individual in the unit set;

[0023] Obtain new individuals based on the best individual in each triangle topological unit and the best individual in the randomly selected unit set; and update the best individual and suboptimal individual through a greedy algorithm;

[0024] Obtain new vertices through the updated optimal and suboptimal individuals;

[0025] The latest individuals are updated according to the new vertices to obtain the optimal number of hidden layer nodes, the optimal initial learning rate, and the optimal L2 regularization coefficient.

[0026] Preferably, the training of the initial thermal error prediction model using multiple sets of temperature data and corresponding thermal errors comprises the following steps:

[0027] Input each set of temperature data into the optimized thermal error prediction model to obtain the corresponding output;

[0028] The difference between the output and the thermal error corresponding to the set of temperature data is back-propagated to train the optimized thermal error prediction model.

[0029] Preferably, before optimizing the model parameters of the initial thermal error prediction model, the thermal error needs to be filtered.

[0030] In a second aspect, the present invention further provides a device for predicting thermal errors of a die bonder, comprising:

[0031] An acquisition module is used to collect multiple sets of temperature data of the die bonding machine and obtain thermal errors corresponding to the multiple sets of temperature data;

[0032] A construction module is used to add multiple LSTM layers and attention layers to the CNN model to obtain an initial thermal error prediction model; the initial thermal error prediction model is trained by using multiple sets of temperature data and corresponding thermal errors to obtain a thermal error prediction model;

[0033] The prediction module is used to input the temperature data to be predicted into the thermal error prediction model, extract the features of the temperature data to be predicted through the CNN layer to obtain the temperature features; extract the time series features of the temperature features through multiple LSTM layers to obtain the global features; and perform weighted summation and mapping of the global features through the attention layer to obtain the thermal error.

[0034] In a third aspect, the present invention further provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned method for predicting thermal errors of a die bonding machine when executing the program.

[0035] In a fourth aspect, the present invention further provides a computer-readable storage medium, wherein the storage medium stores a computer program, and when the computer program is executed by a processor, the above-mentioned method for predicting thermal errors of a die bonding machine is implemented.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] The present invention adds multiple LSTM layers and attention layers to the CNN model to obtain an initial thermal error prediction model, extracts features of the temperature data to be predicted through the CNN layer to obtain temperature features; and extracts time series features of the temperature features through multiple LSTM layers to obtain global features. The present invention combines CNN with LSTM, and can achieve comprehensive extraction of data spatial features and time series features, thereby more accurately describing and understanding the complex patterns of the data. After weighted summation and mapping of global features through the attention layer and incorporating the attention mechanism into the model, it can flexibly respond to different data distributions and patterns. Finally, the optimized thermal error prediction model is trained by multiple sets of temperature data and corresponding positioning errors to obtain a thermal error prediction model. The prediction accuracy of thermal errors is greatly improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0039] Figure 1 It is a flow chart of a method for predicting thermal error of a die bonding machine of the present invention;

[0040] Figure 2 It is a structural schematic diagram of the initial thermal error prediction model of the present invention;

[0041] Figure 3This is a photo of the die bonding machine.

[0042] Figure 4 It is a schematic diagram for comparing the thermal error prediction model of the present invention with the true value. DETAILED DESCRIPTION

[0043] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0044] Reference Figure 1 The present invention provides a method for predicting thermal errors of a die bonder, comprising the following steps:

[0045] Step 1: Collect multiple sets of temperature data from the die bonder and obtain the positioning error corresponding to each set of temperature data.

[0046] Temperature sensors are set at different positions of the X-axis of the die bonder. Each temperature sensor collects a set of temperature data. Each set of temperature data is clustered by a clustering algorithm to obtain temperature sensitive points. Temperature sensitive points refer to the temperature measurement points that have a greater impact on the error among the multiple sets of measured temperature data. The temperature sensitive points are selected to reduce the complexity of the calculation.

[0047] Reference Figure 3 , select the starting point and the end point on the X-axis of the die bonder, the first distance between the two data points is fixed, and a certain speed of the X-axis welding head is given. In theory, the time it takes for the welding head to pass through the two data points is determined. When the welding head is in operation, the timing starts from the position of the starting point. After the theoretical time has passed, the second distance between the welding head and the starting point is obtained.

[0048] Thermal error refers to the difference between the actual position (positioning error) measured and the geometric error. Therefore, when the die bonder is just turned on, the temperature is 0, and the difference between the second distance and the first distance is obtained, which is the geometric error. Under different sensitive temperatures, the difference between the corresponding second distance and the first distance is obtained to obtain the corresponding positioning error. Multiple positioning errors are subtracted from the geometric errors to obtain multiple thermal errors.

[0049] In this embodiment, a camera is provided at the welding head and a scale is provided on one side. The time between each two data points is defined as a photographing cycle. The position of the scale photographed by the camera is the measured value of positioning. The difference between the measured value and the end point is the positioning error.

[0050] The positioning error data is filtered to make it smoother and more suitable for subsequent analysis and modeling.

[0051] Step 2: Add multiple LSTM layers and attention layers to the CNN model to obtain the initial thermal error prediction model, optimize the model parameters of the initial thermal error prediction model to obtain the optimized thermal error prediction model; and train the optimized thermal error prediction model through multiple sets of temperature data and corresponding positioning errors to obtain the thermal error prediction model.

[0052] Reference Figure 2 , establish the initial thermal error prediction model, combine the convolutional neural network CNN with the long short-term memory network LSTM, and introduce the Attention mechanism. The basic structure of the convolutional neural network CNN includes convolution layer, pooling layer and fully connected layer. The calculation logic of the convolution layer in the feature extraction process is:

[0053]

[0054] in, and Respectively represent the weight and bias corresponding to the convolution filter at the (i, j)th position; Represents the feature map of the previous layer; represents the feature map of the current layer; f(·) represents the activation function; M j Represents a collection of feature maps.

[0055] The specific processing process of each model is as follows: CNN-LSTM-Attention, as a deep learning model architecture, is particularly suitable for processing sequence data with spatiotemporal correlation. This structure cleverly combines three different types of neural network layers: convolutional neural network (CNN), long short-term memory network (LSTM) and attention mechanism (Attention), aiming to fully capture the spatial and temporal features in the data, thereby achieving in-depth analysis and accurate prediction of complex sequence data. The structure is as follows. Convolutional neural network (CNN) is an algorithm designed for grid-structured data. It extracts the features of input data layer by layer to achieve automatic learning and classification of data such as images. The core features of CNN are weight sharing and local connection, which helps to reduce the number of model parameters and improve the generalization ability of the model. In addition to the input layer and output layer, the structure of CNN usually consists of multiple convolutional layers, pooling layers and fully-connected layers. Among them, the convolution layer slides on the input data and performs convolution operations through the convolution kernel, which is responsible for extracting the local features of the input data; the pooling layer downsamples the feature map output by the convolution layer to reduce the spatial dimension and computational complexity of the data; the fully connected layer maps the features extracted by the previous layer to the sample label space to achieve classification or regression tasks.

[0056] The model parameters of the initial thermal error prediction model are optimized using the triangular topology clustering optimization algorithm, including the optimal number of hidden layer nodes, the optimal initial learning rate, and the optimal L2 regularization coefficient. The optimization process is divided into the population initialization stage, the triangular topology unit formation stage, the global aggregation stage, and the local aggregation stage.

[0057] Population initialization phase.

[0058] Each vertex in the triangular topological unit represents a search individual, and the mathematical expression generated by each individual is:

[0059]

[0060] Where, X i,1 represents the first search individual in the i-th triangle topological unit, and i is a positive integer between 1 and N / 3, N is the population individual, r0 represents a random number between [0, 1], and are the lower and upper bounds of the variable, and their mathematical expressions can be expressed as:

[0061]

[0062] The stage of forming triangular topological unit.

[0063] In the spherical coordinate system, a specific first vertex is used as the starting point, and a new direction vector with a length of l*f is defined. This direction vector determines the position of the second vertex by converting the spherical coordinates to the ordinary coordinates. Then, this direction vector is rotated counterclockwise by π / 3, and then the position of the third vertex is obtained by the coordinate transformation. The following are the mathematical expressions for determining the positions of these vertices:

[0064]

[0065] Where l represents the size of the triangular topological unit, which is mathematically expressed as:

[0066]

[0067] t represents the current iteration number, T represents the maximum number of iterations. l decreases as the number of iterations increases.

[0068] f(θ) and f(θ+π / 3) represent the direction vectors of the other two edges guided by the first point

[0069]

[0070]

[0071] where θ = [θ1,…,θ D ] and θ j (j=1, ..., D) is a random number between [0, π].

[0072] Each group of triangular topological units is internally aggregated into a fourth vertex. This vertex is formed in a linearly weighted manner to use individual information, defined as:

[0073]

[0074] Where r1, r2 and r3 are random numbers between [0, 1] and r1 + r2 + r3 = 1. Therefore, the fourth search individual is located in each triangular topological unit.

[0075] Global aggregation stage

[0076] At this stage, we focus on collecting information from outstanding individuals in different triangular units, and then construct new and potential solutions. This information interaction occurs between the best individual inside each triangular topological unit and the best individual in the randomly selected unit set. The linear combination method is used to merge the characteristics of these two excellent individuals according to the different weights of each dimension variable. Finally, the new individual is generated at the best connection point between these two high-quality vertices, which is mathematically expressed as:

[0077]

[0078] Where r4 is a random number between [0, 1]. and represents the best individual of unit i and the randomly selected unit at the tth iteration. In addition, the greedy strategy is used to update the best individual, and the mathematical expression is:

[0079]

[0080] Local aggregation stage

[0081] In the current stage, local aggregation operations are performed on triangular topological units. Based on the previous stage, a triangular topological structure is temporarily constructed between the updated optimal or suboptimal individuals and two vertices in the group with good fitness values. However, these topological triangles are not fixed as equilateral triangles, and their shapes will change flexibly according to actual conditions. Based on the difference in motion vectors formed between the optimal and suboptimal individuals, the position of the optimal individual is perturbed and adjusted in the local area, including fine-tuning of the direction and step size. The purpose of this is to re-search each group within a certain local range, so as to achieve more efficient use of each topological triangle unit. The new vertex is calculated as:

[0082]

[0083] α can be calculated as:

[0084]

[0085] The purpose of using the information of suboptimal individuals is to prevent the optimal individual from falling into the local optimal solution. After aggregation, it is necessary to ensure that the guiding point of each temporary triangle unit is optimal inside its unit. In order to guide the convergence process in a more promising direction, the fitness values ​​of the two vertices before and after local mining are compared to decide whether to update the position. If the new individual shows better fitness, that is, better than the original individual, the position will be updated; otherwise, the original position will remain unchanged. The mathematical expression of this process is as follows:

[0086]

[0087] The optimized thermal error prediction model is trained using the preprocessed die bonding machine measurement data, and the trained optimized thermal error prediction model is used as the thermal error prediction model.

[0088] Step 3: Use the thermal error prediction model to predict the thermal error at different temperatures.

[0089] according to Figure 4 It can be seen that the predicted value curve output by the thermal error prediction model is almost consistent with the true value curve, which proves the prediction accuracy of the model of this method.

[0090] Based on the same concept, the present invention also provides a device for predicting thermal errors of a die bonder, comprising an acquisition module, a construction module and a prediction module.

[0091] The acquisition module is used to collect multiple sets of temperature data of the die bonding machine and obtain thermal errors corresponding to the multiple sets of temperature data.

[0092] The building module is used to add multiple LSTM layers and attention layers to the CNN model to obtain an initial thermal error prediction model; the initial thermal error prediction model is trained by multiple sets of temperature data and corresponding thermal errors to obtain a thermal error prediction model.

[0093] The prediction module is used to input the temperature data to be predicted into the thermal error prediction model, extract the features of the temperature data to be predicted through the CNN layer to obtain the temperature features; extract the time series features of the temperature features through multiple LSTM layers to obtain the global features; and perform weighted summation and mapping of the global features through the attention layer to obtain the thermal error.

[0094] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned thermal error prediction method for a die bonding machine when executing the program.

[0095] The present invention also provides a computer-readable storage medium, wherein the storage medium stores a computer program, and when the computer program is executed by a processor, the above-mentioned method for predicting thermal errors of a die bonding machine is implemented.

[0096] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0097] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A method for predicting thermal error of a die bonder, characterized in that: The following steps are involved: Collect multiple sets of temperature data from the die bonder and obtain thermal errors corresponding to the multiple sets of temperature data; Add multiple LSTM layers and attention layers to the CNN model to obtain the initial thermal error prediction model; The initial thermal error prediction model is trained by using multiple sets of temperature data and corresponding thermal errors to obtain a thermal error prediction model; The temperature data to be predicted is input into the thermal error prediction model, and the temperature data to be predicted is subjected to feature extraction through the CNN layer to obtain temperature features; the time series features of the temperature features are extracted through multiple LSTM layers to obtain global features; The global features are weighted summed and mapped through the attention layer to obtain the thermal error.

2. A method for predicting thermal errors of a die bonder according to claim 1, characterized in that: Temperature sensors are set at different positions of the X-axis of the die bonding machine. Each temperature sensor collects a set of temperature data. Each set of temperature data is clustered by a clustering algorithm to obtain multiple sensitive temperatures.

3. A method for predicting thermal errors of a die bonder according to claim 2, characterized in that: The step of obtaining thermal errors corresponding to the multiple sets of temperature data comprises the following steps: Selecting a starting point and an end point on the X-axis of the die bonder, and obtaining a first distance between the starting point and the end point; Obtaining a theoretical time for the X-axis welding head to pass through the first distance according to the running speed of the X-axis welding head; When the welding head is in operation, timing is started from the position of the starting point, and after the theoretical time has passed, the second distance between the welding head and the starting point is obtained; When the die bonding machine is just turned on, a difference between the second distance and the first distance is obtained, where the difference is a geometric error; Under different sensitive temperatures, the difference between the corresponding second distance and the first distance is obtained to obtain the corresponding positioning error; Multiple positioning errors are subtracted from the geometric errors to obtain multiple thermal errors.

4. A method for predicting thermal errors of a die bonder according to claim 1, characterized in that: Before the initial thermal error prediction model is trained by using multiple sets of temperature data and corresponding thermal errors, the model parameters of the initial thermal error prediction model are optimized. The optimization process specifically includes the following steps: Taking different model parameters as an individual, generating individuals with one third of the population; taking multiple individuals as the first vertex, wherein the model parameters include the number of hidden layer nodes, the initial learning rate, and the L2 regularization coefficient; Taking the first vertices as starting points, a direction vector is defined in a spherical coordinate system, and the direction vector is subjected to coordinate transformation to obtain a second vertex; the direction vector is rotated 60 degrees counterclockwise and then subjected to coordinate transformation to obtain a third vertex; the first vertex, the second vertex and the third vertex form a triangular topological unit; forming a fourth vertex in each triangular topological unit in a linearly weighted manner; Obtain the fitness values ​​of all vertices in each triangular topological unit and sort them to obtain the best individual in each triangular topological unit, and randomly select the best individual in the unit set; Obtain new individuals based on the best individual in each triangle topological unit and the best individual in the randomly selected unit set; and update the best individual and suboptimal individual through a greedy algorithm; Obtain new vertices through the updated optimal and suboptimal individuals; The latest individuals are updated according to the new vertices to obtain the optimal number of hidden layer nodes, the optimal initial learning rate, and the optimal L2 regularization coefficient.

5. A method for predicting thermal errors of a die bonder according to claim 1, characterized in that: The initial thermal error prediction model is trained by using multiple sets of temperature data and corresponding thermal errors, comprising the following steps: Input each set of temperature data into the optimized thermal error prediction model to obtain the corresponding output; The difference between the output and the thermal error corresponding to the set of temperature data is back-propagated to train the optimized thermal error prediction model.

6. A method for predicting thermal errors of a die bonder according to claim 5, characterized in that: Before optimizing the model parameters of the initial thermal error prediction model, the thermal error needs to be filtered.

7. A device for predicting thermal errors of a die bonder, characterized in that: include: An acquisition module is used to collect multiple sets of temperature data of the die bonding machine and obtain thermal errors corresponding to the multiple sets of temperature data; A building block is used to add multiple LSTM layers and attention layers to the CNN model to obtain an initial thermal error prediction model; The initial thermal error prediction model is trained by using multiple sets of temperature data and corresponding thermal errors to obtain a thermal error prediction model; The prediction module is used to input the temperature data to be predicted into the thermal error prediction model, extract the features of the temperature data to be predicted through the CNN layer to obtain the temperature features; extract the time series features of the temperature features through multiple LSTM layers to obtain the global features; The global features are weighted summed and mapped through the attention layer to obtain the thermal error.

8. A computer device, characterized in that: The invention comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method for predicting thermal errors of a die bonding machine as described in any one of claims 1 to 6 is implemented.

9. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by a processor, the method for predicting thermal errors of a die bonder according to any one of claims 1 to 6 is implemented.

Images