A machine learning-based uv curing machine running energy consumption analysis method

Abstract

The application relates to the technical field of data processing, and discloses a UV curing machine running energy consumption analysis method based on machine learning, which comprises the following steps: collecting relevant parameters and relevant data of a plurality of circuit boards, recording feature codes of elements on the circuit boards, and recording a plurality of glue coating parameters, curing parameters and energy consumption levels of a UV curing machine in a glue coating and curing process of the circuit boards; constructing a device feature matrix of the circuit boards; obtaining distance consistency of two elements, and then obtaining layout similarity of the two elements; obtaining a plurality of matching element pairs of two circuit boards; quantifying pattern similarity of the two circuit boards; obtaining a plurality of circuit board types; constructing a glue coating parameter prediction model; and constructing a curing energy consumption prediction model to predict energy consumption in the glue coating and curing process of the circuit boards. The application aims to solve the problem that curing process energy consumption of fixed curing parameters for different chip types is obviously different.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of data processing, in particular to a UV curing machine running energy consumption analysis method based on machine learning. BACKGROUND

[0002] The UV curing technology is praised as the "5E" green industrial technology due to its rapid curing, low-temperature operation and low VOC (volatile organic compound) emission. In terms of energy consumption, the energy consumption of the UV curing process is usually only 10%-20% of that of the traditional process, which shows a significant energy-saving advantage. However, "low energy consumption" is a relative concept. The UV curing machine itself, especially its core component UV lamp or LED lamp bead, will generate a large amount of electric energy consumption during operation and produce a lot of heat, which requires an additional cooling system to maintain stable operation of the equipment and ensure the curing quality. Therefore, it is necessary to finely analyze and dynamically optimize the running energy consumption of the UV curing machine.

[0003] The existing UV curing machine energy consumption analysis mostly depends on preset curing parameters for simulation, or determines the corresponding curing parameters according to the current curing materials, and then quantifies the energy consumption of the curing machine under different operating conditions according to the curing process corresponding to the curing machine operating parameters. However, due to the different sizes of the circuit boards to be cured, the UV glue coating conditions on them are different, which leads to obvious differences in the time required for irradiation curing. In addition, different versions and chips require different glue coatings, and the required light intensity during curing is different, which leads to different curing parameters required for the glue on different circuit boards during UV curing, resulting in obvious differences in the energy consumption of the UV curing machine. Ultimately, the method of setting fixed curing parameters according to the existing fixed preset curing parameters or simply according to the glue material leads to obvious differences in the analysis of the energy consumption of the UV curing machine. SUMMARY

[0004] The application provides a UV curing machine running energy consumption analysis method based on machine learning to solve the problem of obvious differences in the energy consumption of the curing process of different versions of chips with fixed curing parameters. The technical scheme adopted is as follows:

[0005] The application provides a UV curing machine running energy consumption analysis method based on machine learning, which comprises the following steps:

[0006] Collecting relevant parameters and data of a plurality of circuit boards, recording the feature codes of each element on the circuit board, and recording a plurality of glue coating parameters, curing parameters and energy consumption levels of the UV curing machine during the glue coating and curing process of each circuit board;

[0007] Based on the distribution and feature coding of electronic components on the circuit board, a device feature matrix of each circuit board is constructed; the performance and distribution of the neighborhood feature coding of two components on two circuit boards in the device feature matrix are analyzed, the distance consistency of the two components is obtained, and the layout similarity of the two components is obtained; based on the layout similarity, the components on the two circuit boards are matched to obtain a plurality of matching component pairs of the two circuit boards;

[0008] According to the feature coding difference and distance consistency between the matching component pairs of the two circuit boards, the pattern similarity of the two circuit boards is quantified; based on the related parameters of all circuit boards, a plurality of circuit board classes are obtained; based on the related data of the same type of circuit board, the pattern similarity between the same type of circuit boards is taken as the attention value, and a glue coating parameter prediction model is constructed; the curing result score of each circuit board is obtained, and the curing parameter prediction model is constructed based on the difference of each glue coating parameter between the circuit boards;

[0009] Based on the curing parameters and energy consumption level of the circuit board coating and curing process, a curing energy consumption prediction model is constructed to predict the energy consumption of the circuit board coating and curing process.

[0010] Optionally, the device feature matrix of each circuit board is constructed by the following specific method:

[0011] For any one circuit board, the circuit board is divided into a plurality of regions as a plurality of initial regions of the circuit board according to the region size; if a plurality of continuous initial regions correspond to an electronic component, the plurality of initial regions corresponding to the electronic component are merged as a small region, and other initial regions are also taken as a small region to obtain a plurality of small regions of the circuit board;

[0012] The feature coding of the electronic components corresponding to each small region is taken as the feature coding of the small region; if there is no electronic component in the corresponding small region, the feature coding of the small region is 0; based on the feature coding of each small region, a device feature matrix of the circuit board is constructed.

[0013] Optionally, the distance consistency of the two components is obtained by the following specific method:

[0014] For the component in the circuit board , the feature coding of the small region in the eight-neighborhood direction in the device feature matrix of the component in the circuit board is obtained; if the feature coding of any neighborhood direction small region is 0, the neighborhood direction is extended until a small region with feature coding not equal to 0 is obtained as the feature coding of the neighborhood direction; the distance between the center of the small region corresponding to the feature coding of each neighborhood direction and the center of the small region corresponding to the component is taken as the distribution distance of each neighborhood direction;

[0015] Acquiring circuit board Intermediate element The feature codes and distribution distances of each neighborhood direction of the two elements in the two circuit boards are obtained, the inverse proportional normalized result of the absolute value of the difference between the distribution distances of the same neighborhood direction of the two elements in the two circuit boards is calculated, and the average of the inverse proportional normalized results corresponding to all neighborhood directions is taken as the distance consistency degree of the two elements in the two circuit boards.

[0016] Optionally, the specific method for obtaining the layout similarity of the two elements comprises the following steps.

[0017] For the circuit board and the circuit board , the number of neighborhood directions with the same feature code of the two elements in the two circuit boards is obtained, and the ratio of the number to the total number of neighborhood directions is taken as the distribution consistency degree of the two elements in the two circuit boards.

[0018] The distribution consistency degree is taken as the weight of the distribution consistency degree, the difference obtained by subtracting the distribution consistency degree from 1 is taken as the weight of the distance consistency degree, the distribution consistency degree and the distance consistency degree of the two elements in the two circuit boards are weighted and summed, and the obtained result is taken as the layout similarity of the two elements in the two circuit boards.

[0019] Optionally, the specific method for obtaining the layout similarity of the two elements comprises the following steps.

[0020] Based on the two-dimensional dynamic time warping algorithm 2D-DTW, each small area in the device feature matrix of the circuit board and the circuit board is matched, the reciprocal of the layout similarity is taken as the distance measure between the corresponding small areas of the two elements in the two circuit boards, and a plurality of matching element pairs of the two circuit boards are obtained.

[0021] Optionally, the specific method for obtaining the layout similarity of the two elements comprises the following steps.

[0022] For any matching element pair of the circuit board and the circuit board , the logical consistency of the feature codes of the two small areas in the matching element pair is obtained, wherein the logical consistency is 1 when the feature codes are completely same, and 0 when the feature codes are different.

[0023] The product of the logical consistency and the distance consistency degree of the corresponding elements of the two small areas is taken as the distribution similarity of the matching element pair, and the ratio of the sum of the distribution similarities of all matching element pairs to the sum of the distance consistency degrees of the corresponding elements of the two small areas in all matching element pairs is taken as the layout similarity of the two circuit boards. ​​

[0024] Optionally, the obtaining a plurality of circuit boards comprises the specific method of:

[0025] Performing density clustering on all the circuit boards, and obtaining a plurality of clustering clusters by using the Euclidean distance between vectors formed by all the relevant parameters of each circuit board as the distance measure, and taking a plurality of circuit boards in the same clustering cluster as the same type of circuit board.

[0026] Optionally, the constructing a glue coating parameter prediction model comprises the specific method of:

[0027] Constructing a glue coating parameter prediction model by using a GNN network, taking the relevant data of a plurality of same type of circuit boards of the target circuit board as the training data set, inputting the relevant data into the glue coating parameter prediction model, the relevant data comprising the EDA principle design drawing and the PCB design drawing of the same type of circuit board, and outputting the multi-class glue coating parameters corresponding to each same type of circuit board, taking the pattern similarity of each same type of circuit board and the target circuit board as the attention value of each same type of circuit board in the training process, and training the glue coating parameter prediction model to obtain the trained glue coating parameter prediction model.

[0028] Inputting the EDA principle design drawing and the PCB design drawing of the target circuit board into the trained glue coating parameter prediction model to obtain the multi-class predicted glue coating parameters of the target circuit board.

[0029] Optionally, the constructing a curing parameter prediction model comprises the specific method of:

[0030] Obtaining the curing score of each circuit board, performing linear normalization on the curing scores of all the circuit boards to obtain the curing result score of each circuit board as the result, obtaining the inverse proportional normalization result of the absolute value difference between any class of glue coating parameter of any circuit board and the predicted glue coating parameter of the target circuit board, and taking the product of the mean value of the inverse proportional normalization result of the absolute value difference of each class of glue coating parameter of the circuit board and the curing result score of the circuit board as the attention value of the circuit board.

[0031] Constructing a curing parameter prediction model by using a long short-term memory network (LSTM) model, taking the glue coating parameters of all the circuit boards under the same type of glue as the training data set, inputting the training data set into the curing parameter prediction model, and outputting the time sequence sequence corresponding to each class of curing parameter in the glue coating process of each circuit board, wherein the length of the time sequence sequence is determined based on the irradiation time length in the curing parameter, to obtain the trained curing parameter prediction model.

[0032] Inputting the predicted glue coating parameters of the target circuit board into the trained curing parameter prediction model to output the predicted irradiation time length of the target circuit board, and further obtaining the predicted time sequence sequence of each class of curing parameter.

[0033] Optionally, the method for constructing the curing energy consumption prediction model comprises the following specific method:

[0034] The curing energy consumption prediction model is constructed based on the CNN network model, time sequence of curing parameters of the UV curing machine in the past period of time is taken as the training data set, input into the curing energy consumption prediction model, and output data is the energy consumption level of the UV curing machine in the past period of time, so that the training completed curing energy consumption prediction model is obtained.

[0035] The beneficial effects of the present application are: the present application constructs a device feature matrix by distributing elements on the circuit board, quantizes the consistent distance and layout similarity of elements in different circuit boards based on the distribution of elements and their feature codes in the device feature matrix, and then matches to obtain the pattern similarity of the circuit board, eliminates the influence of similar models caused by the size difference of the circuit board affected by the function, and filters similar circuit boards by clustering related parameters of the circuit board; according to the related data and gluing parameters of the similar circuit board, and taking the pattern similarity as the attention value to construct the gluing parameter prediction model, the gluing parameter prediction of the target circuit board with similar layout is carried out, and the curing parameter prediction is further carried out on the basis of the gluing parameter, and the influence of the curing result on the gluing and curing result of each circuit board is considered; finally, according to the predicted curing parameters of the target circuit board, the energy consumption of the circuit board to be glued and cured is predicted and analyzed according to the energy consumption level of the curing process in the past, so that the energy consumption prediction distortion caused by the ambiguity of the curing parameters is avoided, and the accuracy of the energy consumption prediction analysis of the UV curing machine in the gluing and curing process of the circuit board is improved. BRIEF DESCRIPTION OF DRAWINGS

[0036] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0037] Figure 1 A machine learning-based UV curing machine running energy consumption analysis method flowchart provided by an embodiment of the present application. DETAILED DESCRIPTION

[0038] The technical solutions in the embodiments of the present application will be described clearly and completely in the following with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0039] Referring to Figure 1 which shows a flow chart of a method for analyzing the energy consumption of a UV curing machine based on machine learning according to an embodiment of the present application, the method comprising the following steps:

[0040] Step S001, collect the relevant parameters and data of a plurality of circuit boards, record the feature codes of the components on the circuit boards, and record the coating parameters, curing parameters and energy consumption levels of the UV curing machine during the coating and curing process of each circuit board.

[0041] The purpose of this embodiment is to predict the energy consumption of the target circuit board to be coated and cured based on the curing parameters and energy consumption levels of the UV curing machine during the coating and curing process of a large number of circuit boards in the past. Therefore, the relevant parameters and data of a large number of circuit boards need to be collected, and the feature codes of the components on each circuit board are recorded to facilitate the subsequent similar analysis of the layout pattern of the circuit boards.

[0042] Specifically, the various curing parameters during the processing of a large number of circuit boards by the UV curing machine are obtained, including various parameters in the UV light source system, cooling system and transmission system, such as UV wavelength, irradiation intensity and irradiation time in the UV light source system, etc. The length of the time sequence is the irradiation time, and the time sequence of various curing parameters of the coating and curing process of each circuit board is obtained, and the energy consumption level of the UV curing machine under different curing parameters is recorded. The coating parameters of the coating and curing process are recorded, including but not limited to glue type, glue coating area, glue coating thickness, etc. The coating parameters corresponding to the same circuit board remain unchanged during the coating and curing process.

[0043] Further, the circuit board to be coated and cured is taken as a target circuit board, and the EDA schematic design diagram and PCB design diagram of the target circuit board and a large number of circuit boards are recorded as the relevant data of the circuit boards. At the same time, a plurality of relevant parameters of the target circuit board and a large number of circuit boards are obtained, including but not limited to input current and voltage, power, and output pin types, etc. The corresponding model, package or supplier number of each component (electronic component) on the target circuit board and a large number of circuit boards is obtained as the feature code of the corresponding component.

[0044] It should be noted that the glue thickness and the coating method required by different chips and other electronic components in the circuit board will be different, resulting in differences in the required light source intensity and time during curing, for example, too long UV irradiation time will cause the glue to harden too much and affect its bonding performance, but if the irradiation time is insufficient, the glue curing effect will be poor, so when curing the circuit board, the relevant parameters during curing need to be adjusted according to the specific situation of the glue coating, to avoid poor curing effect, and due to the adjustment of the curing parameters, it will directly affect the overall energy consumption of the UV curing machine, so in order to accurately evaluate and analyze, the curing parameters of the target circuit board need to be accurately constructed.

[0045] Step S002, based on the distribution of electronic components on the circuit board and the feature code thereof, a device feature matrix of each circuit board is constructed; the performance and distribution of the neighborhood feature codes of two components on two circuit boards in the device feature matrix are analyzed, the distance consistency of the two components is obtained, and then the layout similarity of the two components is obtained; based on the layout similarity, the components on the two circuit boards are matched, and a plurality of matched component pairs of the two circuit boards are obtained.

[0046] It should be noted that since the functionality of the circuit board is composed of electronic components at various locations thereon, the components at various locations on a circuit board are basically certain, and for circuit boards of the same model or type, the components at similar locations are also similar, but there may be differences in the specific component models. Therefore, the device feature matrix is constructed according to the component models at various locations on the circuit board.

[0047] Preferably, in an embodiment of the present application, based on the distribution of electronic components on the circuit board and the feature code thereof, a device feature matrix of each circuit board is constructed, including the following specific method:

[0048] For any one circuit board, the preset region size of the present embodiment is 1mm*1mm, the circuit board is divided into a plurality of regions with the region size as a plurality of initial regions of the circuit board, wherein the initial regions do not overlap, if the remaining size does not meet the region size, it is combined with the previous region as an initial region, wherein the division is first horizontal and then vertical; if a plurality of continuous initial regions correspond to an electronic component, the plurality of initial regions corresponding to the electronic component are combined as a small region, and the other initial regions are also combined as a small region, thereby obtaining a plurality of small regions of the circuit board.

[0049] Furthermore, the feature codes of the electronic components corresponding to each small region are used as the feature codes of that small region. The feature codes are unique and correspond to the model, package, or supplier number of the electronic component. The same feature code corresponds to the same model and size of the component. If there is no electronic component in the corresponding small region, the feature code of that small region is 0. Based on the feature codes of each small region, a device feature matrix is ​​constructed for the circuit board. The rows and columns of the device feature matrix are constructed based on the initial regions. If a small region contains multiple initial regions, the matrix elements corresponding to the multiple initial regions are marked with the feature codes of the small region. The matrix elements corresponding to the initial regions of other small regions are also marked with feature codes, thus obtaining the device feature matrix of the circuit board.

[0050] It should be further explained that in two circuit boards of the same type, the different application ranges lead to certain size adjustments, which increases the distance between two components at the same position on the two circuit boards. However, the layout and model of the two sets of components remain unchanged. Therefore, the spatial distance needs to be scaled by the consistency of the component layout. That is, if the relative positions of the components on two small areas of the two circuit boards are consistent, the proportion of the consistency of their spatial distribution can be reduced to adjust the similarity of the two sets of components.

[0051] Preferably, in one embodiment of the present invention, the method of analyzing the performance and distribution of the neighborhood feature codes of two components on two circuit boards in the device feature matrix to obtain the distance consistency between the two components, and then obtaining the layout similarity between the two components, includes the following specific methods:

[0052] For circuit boards medium components Obtain components On the circuit board The feature codes of small regions in the eight neighborhood directions of the device feature matrix are determined. If the feature code of a small region in any neighborhood direction is 0, the extension is continued along that neighborhood direction until a small region with a non-zero feature code is obtained, which is then used as the feature code for that neighborhood direction. Since a small region may contain multiple initial regions, this is represented by multiple matrix elements with the same feature code in the device feature matrix. Neighborhood acquisition is performed by starting with the outermost matrix element along the corresponding neighborhood direction. Simultaneously, the center of the small region corresponding to the feature code of each neighborhood direction and the element are obtained. The distance between the centers of the corresponding small regions is used as the distribution distance in each neighborhood direction; that is, if a small region consists of multiple initial regions, the distance between the centers is obtained; obtain the circuit board. medium components The feature codes and distribution distances of each neighborhood direction are used to calculate the inverse proportional normalization result of the absolute value of the difference between the distribution distances of the two components in the same neighborhood direction in the two circuit boards. The mean of the inverse proportional normalization results corresponding to all neighborhood directions is used as the distance consistency between the two components in the two circuit boards.

[0053] Furthermore, for circuit boards medium components and circuit boards medium components The number of identical neighborhood directions of two components on two circuit boards is obtained by acquiring the feature encoding of the same neighborhood direction. The ratio of the number to the total number of neighborhood directions (8) is used as the distribution consistency of the two components on the two circuit boards. The distribution consistency is used as the weight of the distribution consistency, and the difference obtained by subtracting the distribution consistency from 1 is used as the weight of the distance consistency. The distribution consistency and distance consistency of the two components on the two circuit boards are weighted and summed, and the result is used as the layout similarity of the two components on the two circuit boards.

[0054] It should be further explained that when circuit boards are similar, the layout and model of various components on their circuit boards are approximately the same, resulting in similar feature matrices of the corresponding components on the boards. Therefore, the similarity between two circuit boards can be evaluated based on the similarity of the feature matrices of the components on the boards. In order to construct the pattern similarity between the two circuit boards, it is first necessary to match the components in the two circuit boards based on the layout similarity.

[0055] Preferably, in one embodiment of the present invention, matching components on two circuit boards based on the layout similarity to obtain several pairs of matching components on the two circuit boards includes the following specific method:

[0056] Based on the 2D Dynamic Time Warping (2D-DTW) algorithm, the circuit board... With circuit board Matching is performed on each small region in the device feature matrix. The reciprocal of the layout similarity is used as the distance metric between the corresponding small regions of the two components in the two circuit boards. Matching is also performed based on the small regions. The row and column of the small region in the device feature matrix are based on the matrix element corresponding to its center, thereby obtaining several matching component pairs between the two circuit boards. It should be noted that the small regions with feature encoding of 0 do not participate in the distance consistency and layout similarity calculation, and they are not processed in the matching process. Therefore, the matched component pairs are all the small regions corresponding to the components.

[0057] At this point, several pairs of matching components for the two circuit boards have been obtained.

[0058] Step S003, quantifying the pattern similarity of the two circuit boards according to the feature code difference and distance consistency between the matching element pairs of the two circuit boards; clustering all circuit boards based on their related parameters to obtain a plurality of categories of circuit boards; constructing a glue coating parameter prediction model based on the related data of the same category of circuit boards, taking the pattern similarity between the same category of circuit boards as the attention value; obtaining the curing result score of each circuit board, and constructing a curing parameter prediction model combined with the difference of each glue coating parameter between the circuit boards.

[0059] Preferably, in an embodiment of the present application, the pattern similarity of the two circuit boards is quantified according to the feature code difference and distance consistency between the matching element pairs of the two circuit boards, which includes the following specific method:

[0060] For any matching element pair in the circuit board and the circuit board , obtain the logical consistency of the feature codes of the two small areas in the matching element pair, wherein the logical consistency is 1 if the feature codes are completely the same, and 0 if they are not the same; take the product of the logical consistency and the distance consistency of the corresponding elements of the two small areas as the distribution similarity of the matching element pair; take the ratio of the sum of the distribution similarities of all matching element pairs to the sum of the distance consistencies of the corresponding elements of all matching element pairs as the pattern similarity of the two circuit boards.

[0061] It should be further pointed out that before UV curing, there are many components in the circuit board, and the areas of the components that need to be cured are similar in the same type of circuit board. For the same type of circuit board or similar circuit board, the layout and related model of the components are basically consistent, so the circuit boards similar to the target circuit board can be selected according to the consistency of the component layout and model on the plurality of circuit boards.

[0062] Preferably, in an embodiment of the present application, all circuit boards are clustered based on their related parameters to obtain a plurality of categories of circuit boards, which includes the following specific method:

[0063] DBSCAN clustering is performed on all circuit boards (including the target circuit board), the Euclidean distance between the vectors composed of all related parameters of each circuit board is used for distance measurement, the preset minimum number of clustering points minPts is 10, the preset radius eps is determined by the distance graph method, the value of k is the preset minPts when the preset radius is determined by the k-distance graph method, a plurality of clustering clusters are obtained, a plurality of circuit boards in the same clustering cluster are taken as the same category of circuit boards, wherein the DBSCAN clustering is a prior art, and the specific clustering process and parameters will not be described herein; then a plurality of same category circuit boards corresponding to the target circuit board are obtained.

[0064] ​Further need to be explained is that when the circuit boards are the same or are under the same model, the components and devices that need to be glued and cured are consistent, and for the same components and devices, in order to ensure that the working state is good, the selection of the glue and the related parameters such as the gluing thickness are similar, and then the gluing parameters of the target circuit board can be predicted according to the gluing parameters and the selection of the same circuit board.

[0065] Preferably, in an embodiment of the present application, based on the related data of the same circuit board, the pattern similarity between the same circuit boards is taken as the attention value, and a gluing parameter prediction model is constructed, including the following specific method:

[0066] The GNN network is used to construct the gluing parameter prediction model, the related data of several same circuit boards of the target circuit board are taken as the training data set and input into the gluing parameter prediction model, the related data includes the EDA principle design diagram and the PCB design diagram of the same circuit board, and the output is the multi-class gluing parameters corresponding to each same circuit board, including but not limited to the glue type, the gluing area, the gluing thickness, etc., and at the same time, the pattern similarity between each same circuit board and the target circuit board is taken as the attention value of each same circuit board in the training process, so as to train the gluing parameter prediction model, and obtain the trained gluing parameter prediction model; the EDA principle design diagram and the PCB design diagram of the target circuit board are input into the trained gluing parameter prediction model, and the multi-class predicted gluing parameters of the target circuit board are obtained.

[0067] Further need to be explained is that because different glues need different UV light conditions when curing, and with the change of the gluing thickness and the gluing area, the related parameters such as the irradiation intensity and the time length of the UV light also need to be modified accordingly, so the curing parameters of the same model glue can be used to predict the curing parameters of the glue on the target circuit board.

[0068] Preferably, in an embodiment of the present application, the curing result scores of each circuit board are obtained, and a curing parameter prediction model is constructed in combination with the differences between each gluing parameter of the circuit boards, including the following specific method:

[0069] The curing results of each circuit board are scored by a UV curing industry expert, and the curing scores of each circuit board are obtained. The curing scores of all circuit boards are linearly normalized, and the obtained results are taken as the curing result scores of each circuit board. The inverse proportional normalized results of the absolute value of the difference between any one gluing parameter of any one circuit board and the predicted gluing parameter of the target circuit board are obtained. The product of the mean value of the inverse proportional normalized results of the absolute value of the difference of each gluing parameter of the circuit board and the curing result score of the circuit board is taken as the attention value of the circuit board.

[0070] Further, based on the RNN network, a long short-term memory (LSTM) model is used to construct the curing parameter prediction model. The coating parameters of all the circuit boards under the same type of glue (the curing conditions are the same, and the target circuit board is also the circuit board for the same type of glue) are used as the training data set, input into the curing parameter prediction model, and the output is the time sequence sequence of each type of curing parameter of each circuit board in the coating process. The length of the time sequence sequence is determined based on the irradiation time length in the curing parameter, and the trained curing parameter prediction model is obtained.

[0071] It should be noted that in the training process, the irradiation time length of each circuit board is known, and the length of the time sequence is determined. At the same time, other curing parameters at each time in the coating process of each circuit board, such as the UV wavelength and irradiation intensity in the UV light source system, are known, and the time sequence sequence corresponding to each type of curing parameter is output.

[0072] Further, the predicted coating parameters of the target circuit board are input into the trained curing parameter prediction model, and the predicted irradiation time length of the target circuit board is output, and the predicted time sequence sequence of each type of curing parameter is obtained.

[0073] At this point, the coating parameter prediction model and the curing parameter prediction model for the target circuit board are constructed.

[0074] Step S004, based on the curing parameters and energy consumption levels of the circuit board coating and curing process, a curing energy consumption prediction model is constructed to predict the energy consumption of the circuit board coating and curing process.

[0075] It should be noted that after obtaining the predicted time sequence sequence of the curing parameters of the coating on the target circuit board in the UV curing machine coating and curing process, the specific energy consumption of the UV curing machine at each time under the corresponding parameters is summed up to obtain the specific possible energy consumption level of the target circuit board coating and curing.

[0076] Specifically, based on the CNN network model, the curing energy consumption prediction model is constructed, and the time sequence sequence of the curing parameters of the UV curing machine in the past period of time is used as the training data set. In this embodiment, the period of time is preset to 1 minute, which is input into the curing energy consumption prediction model. The output data is the energy consumption level of the UV curing machine in the past period of time, and the trained curing energy consumption prediction model is obtained.

[0077] Further, the prediction time sequence of each type of curing parameter of the target circuit board is segmented, the continuous time points with the same each type of curing parameter are formed into a segment, and a plurality of segments corresponding to the prediction time sequence of each type of curing parameter are obtained, and the each type of curing parameter of all time points in a segment is the same; the each type of curing parameter of any segment is input into the trained curing energy consumption prediction model, and the prediction energy consumption level of the segment is output, and then the prediction energy consumption level of each segment in the glue coating and curing process of the target circuit board is obtained, and the sum of all prediction energy consumption levels is used as the prediction energy consumption of the glue coating and curing process of the target circuit board, and the energy consumption prediction and analysis of the UV curing machine for the glue coating and curing process of the circuit board is realized.

[0078] It should be noted that the embodiment adopts a model to present an inverse proportional relationship and normalization processing, as the input of the model, represents an exponential function with a natural constant as the base, and the implementer can set the inverse proportional function and the normalization function according to the actual situation.

[0079] Thus, the embodiment is completed.

[0080] The above only describes the preferred embodiment of the present application and is not used to limit the present application, and any modification, equivalent replacement, improvement, etc. within the principle of the present application should be included in the protection scope of the present application.

Claims

1. A method for analyzing energy consumption of a UV curing machine based on machine learning, characterized in that, The method comprises the following steps: Collecting relevant parameters and data of a plurality of circuit boards, recording feature codes of each component on the circuit boards, and recording a plurality of coating parameters, curing parameters and energy consumption levels of the UV curing machine in the gluing and curing process of each circuit board; Based on the distribution and feature codes of electronic components on the circuit board, a component feature matrix of each circuit board is constructed; the consistency of the distance between two components in the component feature matrix is analyzed based on the neighborhood feature codes of the two components on the two circuit boards, and the layout similarity of the two components is obtained; based on the layout similarity, the components on the two circuit boards are matched to obtain a plurality of matched component pairs of the two circuit boards; According to the feature code difference and distance consistency between the matched component pairs of the two circuit boards, the pattern similarity of the two circuit boards is quantified; all circuit boards are clustered based on their relevant parameters to obtain a plurality of circuit board classes; based on the relevant data of the same circuit board class, the pattern similarity between the same circuit board classes is taken as the attention value to construct a coating parameter prediction model; the curing result score of each circuit board is obtained, and a curing parameter prediction model is constructed based on the difference of each coating parameter between the circuit boards; Based on the curing parameters and energy consumption levels of the gluing and curing process of the circuit board, a curing energy consumption prediction model is constructed to predict the energy consumption of the gluing and curing process of the circuit board; The specific method for constructing the component feature matrix of each circuit board comprises: For any one circuit board, the circuit board is divided into a plurality of regions as a plurality of initial regions of the circuit board according to the region size; if a plurality of continuous initial regions correspond to one electronic component, the plurality of initial regions corresponding to the electronic component are merged as one small region, and other initial regions are also taken as one small region to obtain a plurality of small regions of the circuit board; The feature codes of the electronic components corresponding to each small region are taken as the feature codes of the small region; if there is no electronic component in the corresponding small region, the feature code of the small region is 0; based on the feature codes of the small regions, a component feature matrix of the circuit board is constructed; The specific method for obtaining the distance consistency of the two components comprises: For component i in circuit board a, the feature codes of the small regions in the eight neighborhood directions of component i in the component feature matrix of circuit board a are obtained; if the feature code of any neighborhood direction small region is 0, the neighborhood direction is extended until a small region with a feature code other than 0 is obtained as the feature code of the neighborhood direction; the distance between the center of the small region corresponding to the feature code of each neighborhood direction and the center of the small region corresponding to component i is taken as the distribution distance of each neighborhood direction; The feature codes and distribution distances of each neighborhood direction of component j in circuit board b are obtained, the absolute value of the difference between the distribution distances of the same neighborhood direction of the two components in the two circuit boards is calculated, the inverse proportional normalization result of the absolute value is obtained, and the average value of the inverse proportional normalization results of all neighborhood directions is taken as the distance consistency of the two components in the two circuit boards.

2. The method of claim 1, wherein the method is based on machine learning. The specific method for obtaining the layout similarity of the two components comprises: For the element i in the circuit board a and the element j in the circuit board b, the feature codes of the same neighborhood direction of the two elements in the two circuit boards are obtained, the number of the same neighborhood direction is obtained, and the ratio of the number to the total number of neighborhood directions is taken as the distribution consistency of the two elements in the two circuit boards; The distribution consistency of the two elements in the two circuit boards is weighted by taking the distribution consistency as the weight of the distribution consistency, and the difference obtained by subtracting the distribution consistency from 1 is taken as the weight of the distance consistency, and the weighted sum of the distribution consistency and the distance consistency of the two elements in the two circuit boards is taken as the layout similarity of the two elements in the two circuit boards. 3.The method of claim 1, wherein, The specific method for obtaining the plurality of matching element pairs of the two circuit boards comprises: Based on the two-dimensional dynamic time warping algorithm 2D-DTW, the device feature matrix of the circuit board a and the circuit board b is matched, the reciprocal of the layout similarity is taken as the distance measure between the corresponding small regions of the two element pairs in the two circuit boards, and the plurality of matching element pairs of the two circuit boards is obtained.

4. The method of claim 1, wherein the method is based on machine learning. The specific method for obtaining the pattern similarity of the two circuit boards comprises: For any matching element pair in the circuit board a and the circuit board b, the logical consistency of the feature codes of the two small regions in the matching element pair is obtained, wherein the logical consistency is 1 when the feature codes are completely same, and the logical consistency is 0 when the feature codes are different; The product of the logical consistency and the distance consistency of the corresponding elements of the two small regions is taken as the distribution similarity of the matching element pair, and the ratio of the sum of the distribution similarities of all matching element pairs to the sum of the distance consistencies of the corresponding elements of the two small regions in all matching element pairs is taken as the pattern similarity of the two circuit boards.

5. The method of claim 1, wherein the method is based on machine learning. The specific method for obtaining the plurality of circuit boards of the same type comprises: Density clustering is performed on all circuit boards, the Euclidean distance between the vectors formed by all related parameters of the circuit boards is used as the distance measure, and a plurality of clustering clusters is obtained, and the plurality of circuit boards in the same clustering cluster are taken as the circuit boards of the same type.

6. The method of claim 1, wherein the method is based on machine learning. The specific method for constructing the glue coating parameter prediction model comprises: A glue coating parameter prediction model is constructed by using a GNN network, related data of a plurality of circuit boards of the same type of a target circuit board is taken as a training data set, the related data includes EDA principle design drawings and PCB design drawings of the circuit boards of the same type, and the output is a plurality of glue coating parameters corresponding to each circuit board of the same type, the pattern similarity of each circuit board of the same type and the target circuit board is taken as the attention value of each circuit board of the same type in the training process, and the glue coating parameter prediction model is trained to obtain a trained glue coating parameter prediction model; The EDA principle design drawings and the PCB design drawings of the target circuit board are input into the trained glue coating parameter prediction model to obtain a plurality of predicted glue coating parameters of the target circuit board.

7. The method of claim 6, wherein the method is based on machine learning. The specific method for constructing the curing parameter prediction model comprises: The curing score of each circuit board is obtained, the curing scores of all circuit boards are linearly normalized, and the obtained result is taken as the curing result score of each circuit board; the inverse proportional normalized result of the absolute value of the difference between any type of glue coating parameter of any circuit board and the predicted glue coating parameter of the target circuit board is obtained, and the product of the mean value of the inverse proportional normalized results of the absolute values of the differences of various types of glue coating parameters of the circuit board and the curing result score of the circuit board is taken as the attention value of the circuit board; A long short-term memory network LSTM model is used to construct a curing parameter prediction model, all glue coating parameters of circuit boards under the same type of glue are taken as training data sets, input into the curing parameter prediction model, and the output is a time sequence corresponding to each type of curing parameter of each circuit board in the glue coating process, wherein the length of the time sequence is determined based on the irradiation time length in the curing parameter, and a trained curing parameter prediction model is obtained; The predicted glue coating parameters of the target circuit board are input into the trained curing parameter prediction model, the predicted irradiation time length of the target circuit board is output, and the predicted time sequence of each type of curing parameter is obtained.

8. The method of claim 7, wherein the method is based on machine learning. The specific method for constructing the curing energy consumption prediction model includes: A CNN network model is used to construct a curing energy consumption prediction model, the time sequence of the curing parameters of the UV curing machine in the past period of time is taken as the training data set, input into the curing energy consumption prediction model, and the output data is the energy consumption level of the UV curing machine in the past period of time, and a trained curing energy consumption prediction model is obtained.

Images