CN115261963A - Method for improving deep plating capability of PCB - Google Patents
Method for improving deep plating capability of PCB Download PDFInfo
- Publication number
- CN115261963A CN115261963A CN202211177716.XA CN202211177716A CN115261963A CN 115261963 A CN115261963 A CN 115261963A CN 202211177716 A CN202211177716 A CN 202211177716A CN 115261963 A CN115261963 A CN 115261963A
- Authority
- CN
- China
- Prior art keywords
- vector
- pcb
- dimension
- neural network
- loss function
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- C—CHEMISTRY; METALLURGY
- C25—ELECTROLYTIC OR ELECTROPHORETIC PROCESSES; APPARATUS THEREFOR
- C25D—PROCESSES FOR THE ELECTROLYTIC OR ELECTROPHORETIC PRODUCTION OF COATINGS; ELECTROFORMING; APPARATUS THEREFOR
- C25D21/00—Processes for servicing or operating cells for electrolytic coating
- C25D21/12—Process control or regulation
-
- C—CHEMISTRY; METALLURGY
- C25—ELECTROLYTIC OR ELECTROPHORETIC PROCESSES; APPARATUS THEREFOR
- C25D—PROCESSES FOR THE ELECTROLYTIC OR ELECTROPHORETIC PRODUCTION OF COATINGS; ELECTROFORMING; APPARATUS THEREFOR
- C25D7/00—Electroplating characterised by the article coated
- C25D7/12—Semiconductors
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Chemical & Material Sciences (AREA)
- Theoretical Computer Science (AREA)
- Chemical Kinetics & Catalysis (AREA)
- Electrochemistry (AREA)
- Materials Engineering (AREA)
- Metallurgy (AREA)
- Organic Chemistry (AREA)
- Physics & Mathematics (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- Artificial Intelligence (AREA)
- Health & Medical Sciences (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Automation & Control Theory (AREA)
- Electroplating Methods And Accessories (AREA)
Abstract
The invention discloses a method for improving the deep plating capacity of a PCB (printed circuit board), belonging to the technical field of circuit boards; the method comprises the following steps: acquiring control parameters of the electroplating process of the PCB for multiple times, and acquiring the deep plating capacity of the PCB in each electroplating process; acquiring a final dimensionality reduction matrix when the first neural network converges; reducing the dimension of each multi-dimensional vector according to the final dimension reduction matrix to obtain a plurality of second dimension reduction vectors; acquiring a trained second neural network; and continuously updating the magnitude of the numerical value in the multi-dimensional vector based on the objective function, and training the second neural network again to obtain the PCB electroplating process control parameter when the deep plating capacity of the PCB is maximum. According to the invention, the optimal electroplating process control parameter data is obtained by using the accurate deep plating capability second neural network, the electroplating process is controlled by using the optimal electroplating process control parameter data, and the deep plating capability of the PCB is effectively improved.
Description
Technical Field
The invention relates to the technical field of circuit boards, in particular to a method for improving the deep plating capacity of a PCB.
Background
With the development of social economy and the improvement of the industrialization level, the usage amount of the PCB is also increased. The electroplating quality of the PCB is an important factor influencing the circuit bearing capacity of the PCB. One of the important factors affecting the electroplating quality of the PCB is the deep plating capability of the PCB. Therefore, the phenomenon that the coating on the outer surface of the orifice is thicker and the coating on the inner wall of the hole is thinner easily occurs in the electroplating process of the PCB under the influence of various factors, and the phenomenon is caused by insufficient deep plating capability. The main factors influencing the deep-plating capability are process data such as current density, liquid medicine concentration, light agent composition, vibration equipment information, jet flow equipment information and the like, and corresponding process control parameters when the optimal deep-plating capability is determined by fitting the relationship between the information and the deep-plating capability. Some electroplating processing data of historical production need to be obtained firstly in order to fit the relation network, but the electroplating processing data are obtained by measurement in actual production or experimental environment, human factors exist in the collection of the process control parameter data, for example, the positions, angles and the like of placing a PCB (printed circuit board), a jet flow machine and the like cannot be guaranteed to be the same when data are collected every time, or the operation flow of each electroplating process cannot be kept the same, or other factors, such as liquid medicine residue caused by that equipment is not cleaned, instability caused by long-time operation of the equipment and the like cannot be considered, so that the collected process control parameters are noisy under the condition, and the process parameters for controlling the deep plating capacity are influenced finally.
Disclosure of Invention
The invention provides a method for improving the deep plating capability of a PCB (printed circuit board). The method utilizes a second neural network with accurate deep plating capability to obtain optimal electroplating process control parameter data, utilizes the optimal electroplating process control parameter data to control an electroplating process, and effectively improves the deep plating capability of the PCB.
The invention aims to provide a method for improving the deep plating capability of a PCB, which comprises the following steps:
acquiring control parameters of the electroplating process of the PCB for multiple times, and acquiring the deep plating capacity of the PCB in each electroplating process; the control parameters of the PCB electroplating process each time form multidimensional vectors;
discretizing according to the deep plating capability of the PCB in each electroplating process to obtain a deep plating capability grade vector;
reducing the dimension of the multi-dimensional vector according to a preset initial dimension reduction matrix to obtain a first dimension reduction vector; wherein the dimension of the first dimension reduction vector is N dimension;
constructing a first neural network according to the first dimension reduction vector as an input and the deep plating capability grade vector as an output; respectively taking the front N/2 dimension and the rear N/2 dimension in the first dimension reduction vector as the input of a first neural network to obtain a first loss function; training a first neural network based on a first loss function, and acquiring a final dimensionality reduction matrix when the first neural network is converged;
reducing the dimension of each multi-dimensional vector according to the final dimension reduction matrix to obtain a plurality of second dimension reduction vectors;
in the process of obtaining the second dimension-reducing vector according to the final dimension-reducing matrix, the noise information in each multi-dimensional vector is separated, so that the front N/2 dimension of each obtained second dimension-reducing vector contains the noise information of the corresponding multi-dimensional vector, and the rear N/2 dimension does not contain the noise information of the corresponding multi-dimensional vector;
acquiring the abnormal degree of any numerical value in the multi-dimensional vectors corresponding to each second dimension-reducing vector according to the numerical value in the final dimension-reducing matrix and the front N/2-dimensional numerical value of each second dimension-reducing vector;
constructing a second neural network according to the multidimensional vector as input and the deep plating capability as output;
acquiring a regularization coefficient weight value of input layer neuron parameters in a second neural network according to the abnormal degree of any numerical value in each multi-dimensional vector; acquiring a comprehensive loss function according to the regularization coefficient weight value and the neuron parameter value of the input layer in the second neural network; training a second neural network based on a data set consisting of a comprehensive loss function, a multi-dimensional vector and a deep plating capability to obtain the trained second neural network;
reducing the dimension of a multidimensional vector formed by electroplating process control parameters of any PCB through a final dimension reduction matrix to obtain a third dimension reduction vector, and obtaining a target function by taking the last N/2 dimension of the third dimension reduction vector as the input of a second neural network; and continuously updating the magnitude of the numerical value in the multi-dimensional vector based on the objective function, and training the second neural network again to obtain the PCB electroplating process control parameter when the deep plating capacity of the PCB is maximum.
In an embodiment, in the process of training the first neural network, a data set composed of multidimensional vectors and deep plating capability grade vectors is adopted, neural network parameters are continuously updated in the training process, an initial dimensionality reduction matrix corresponding to the first dimensionality reduction vector is continuously updated and obtained until convergence, and the initial dimensionality reduction matrix when the first neural network converges is obtained to serve as a final dimensionality reduction matrix.
In one embodiment, the process of obtaining the comprehensive loss function is to obtain a regularization loss function according to a regularization coefficient weight value and a neuron parameter value of an input layer in a second neural network; and obtaining the normalized loss function and the original mean square error loss function of the second neural network.
In one embodiment, in the process of obtaining the control parameter of the plating process of the PCB when the deep plating capability of the PCB is maximized, based on the objective function, while keeping the parameter in the second neural network unchanged, the magnitude of the numerical value in the multidimensional vector is continuously updated by adopting a gradient descent method to train the second neural network again until convergence, and the multidimensional vector in the final convergence is obtained, that is, the control parameter of the plating process of the PCB when the multidimensional vector corresponds to the maximization of the deep plating capability of the PCB is obtained.
In one embodiment, the first loss function is obtained according to the following steps:
inputting the last N/2 dimension of the first dimensionality reduction vector into a first neural network to obtain a first prediction deep plating capacity grade vector;
acquiring a second loss function according to the first prediction deep plating capability grade vector;
inputting the first N/2 dimension of the first dimension reduction vector into a first neural network to obtain a second prediction deep plating capability grade vector;
acquiring a third loss function according to the second prediction deep plating capability grade vector;
and obtaining a first loss function according to the second loss function and the third loss function.
In one embodiment, the second loss function is as follows:
indicates that the plating depth of the kth electroplating process is graded asThe throwing power level vector of;
indicates that the plating depth of the kth electroplating process is graded asThe first predictive throwing power level vector of (1).
In one embodiment, the third loss function is as follows:
indicates that the plating depth of the kth electroplating process is graded asThe second predictive throwing power level vector of (1).
In one embodiment, the deep plating capability of the PCB in each electroplating process is obtained according to the following steps:
the plating thickness of PCB diaphragm orifice inner wall and the epitaxial plating thickness in PCB diaphragm orifice are detected through cladding material thickness detector, obtain the deep-plating ability according to the ratio of the plating thickness in diaphragm orifice inner wall and the epitaxial plating thickness in diaphragm orifice.
In one embodiment, the throwing power level vector is obtained according to the following steps:
discretizing the deep plating capacity of the PCB in each electroplating process into a plurality of quality grades of the deep plating capacity;
and acquiring the quality grade of the deep plating capability corresponding to the deep plating capability of the PCB in each electroplating process, and performing one-hot coding on the quality grade of the deep plating capability to acquire a deep plating capability grade vector corresponding to each quality grade of the deep plating capability.
The invention has the beneficial effects that:
the method monitors the dimension reduction of each process control parameter data by designing a loss function, separates the accurate separable dimension data of the dimension-reduced data from the noise inseparable dimension data, and determines the abnormal degree of each process control parameter data by analyzing the information content of each data in the noise inseparable dimension data. And determining a self-adaptive regularization loss function through the abnormal degree, and performing supervised network training by using the loss function to prevent the occurrence of a data abnormal overfitting phenomenon and improve the accuracy of the deep plating capability second neural network. And then, the second neural network with the accurate deep plating capability is used for acquiring the optimal electroplating process control parameter data, and the optimal electroplating process control parameter data is used for controlling the electroplating process, so that the deep plating capability of the PCB is effectively improved.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the embodiments or the description of the prior art will be briefly described below, it is obvious that the drawings in the following description are only some embodiments of the present invention, and for those skilled in the art, other drawings can be obtained according to the drawings without creative efforts.
FIG. 1 is a flow chart illustrating the general steps of an embodiment of the method for improving the deep plating capability of a PCB board of the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
The invention monitors the dimension reduction of each process control parameter data by designing a loss function, so that the accurate separable dimension data of the dimension-reduced data is distinguished from the noise inseparable dimension data, and the abnormal degree of each process control parameter data is determined by analyzing the information content of each data in the noise inseparable dimension data. And determining a self-adaptive regularization loss function through the abnormal degree, and performing supervised network training by using the loss function to prevent the occurrence of a data abnormal overfitting phenomenon and improve the accuracy of the deep plating capability second neural network. And then, the second neural network with the accurate deep plating capability is used for acquiring the optimal electroplating process control parameter data, and the optimal electroplating process control parameter data is used for controlling the electroplating process, so that the deep plating capability of the PCB is effectively improved.
The present invention is directed to the following scenarios: the electroplating data of the PCB in the production process are obtained, the data are used for training the second neural network with the deep plating capability, the network training is controlled by controlling the overfitting influence of the abnormal information on the second neural network with the deep plating capability, an accurate second neural network with the deep plating capability is further obtained, and the optimal electroplating control parameters are obtained by using the second neural network with the deep plating capability.
The invention provides a method for improving the deep plating capacity of a PCB (printed circuit board), which is shown in figure 1 and comprises the following steps:
s1, acquiring control parameters of a multi-time PCB electroplating process, and acquiring the deep plating capacity of the PCB in each electroplating process; the control parameters of the PCB electroplating process each time form multidimensional vectors;
in the present embodiment, each time is obtained assuming that the plating process was performed K times in the historyThe process control parameters in the secondary electroplating process are as follows: the PCB electroplating process has the following control parameters: the concentration of the liquid medicine, the current density, the composition of the light agent, vibration information (vibration frequency and amplitude) and jet flow information (jet flow speed) are required to be collected. For the convenience of the following description, the present embodiment forms a vector by the process control parameters used in the Kth electroplating processThe vector has an H dimension;
in addition, the deep plating capability of the PCB in each electroplating process is obtained, and for the convenience of subsequent description, the deep plating capability of the PCB measured in the kth electroplating process is recorded as(ii) a Wherein K =1,2 \8230, K; wherein K represents the number of plating process operations.
The method for acquiring the deep plating capacity of the PCB comprises the following steps: detecting the thickness of a coating on the inner wall of a PCB hole and the thickness of a coating on the extension of the PCB hole by a coating thickness detector; and taking the ratio of the thickness of the coating on the inner wall of the hole to the thickness of the coating on the extension of the hole as an evaluation index of the deep plating capability.
Constructing the acquired data into a first data set: acquiring electroplating process control parameter data and deep plating capacity data in K times of historical processing processes of the PCB:as a first data set.
S2, discretizing according to the deep plating capability of the PCB in each electroplating process to obtain a deep plating capability grade vector;
the deep plating capacity grade vector is obtained according to the following steps:
discretizing the deep plating capacity of the PCB in each electroplating process into a plurality of quality grades of the deep plating capacity;
and acquiring the quality grade of the deep plating capability corresponding to the deep plating capability of the PCB in each electroplating process, and performing one-hot coding on the quality grade of the deep plating capability to acquire a deep plating capability grade vector corresponding to each quality grade of the deep plating capability.
In this embodiment, discretization processing is performed on the acquired data to form a second data set;
discretizing the deep plating capability of the PCB in each electroplating process into a plurality of quality grades of the deep plating capability, which comprises the following steps:
dividing the deep plating capacity into a plurality of discrete levels from small to large by 0.1 as the interval of the deep plating capacityRespectively representing a deep-plating ability grade 1, a deep-plating ability grade 2, \ 8230and a deep-plating ability grade 10. To this end, the throwing power is divided into 10 levels or into 10 categories;
obtainingCorresponding deep plating ability grade, one-hot coding is carried out on the deep plating ability grade, and then a 10-dimensional vector is obtained(ii) a The 10-dimensional vector is the deep plating capacity grade vector;
To this end, the data after the discrete processing is configured into a second data set: acquiring electroplating process control parameter data and deep plating capability grade vector data in K times of historical processing processes of the PCB:as a second data set.
S3, reducing the dimension of the multi-dimensional vector according to a preset initial dimension reduction matrix to obtain a first dimension reduction vector; the dimensionality of the first dimensionality reduction vector is N-dimensionality;
constructing a first neural network according to the first dimension reduction vector as an input and the deep plating capability grade vector as an output; respectively taking the front N/2 dimension and the rear N/2 dimension in the first dimension reduction vector as the input of a first neural network to obtain a first loss function; training a first neural network based on a first loss function, and acquiring a final dimensionality reduction matrix when the first neural network converges;
it should be noted that, for perfect data, any dimension of the data can be classified well, and when some dimensions of data in the data are polluted by noise, the data of some dimensions cannot be separated well, so that a dimension reduction control loss function is constructed by using the feature, and the accurate data dimension and the noise data dimension of the data after dimension reduction are separated. For one data in the data setFirstly, toThe dimension reduction processing is carried out, and the data after the dimension reduction is expected in the embodiment can be used forThe noise contained therein is separated out.
The specific first loss function is obtained according to the following steps:
inputting the last N/2 dimension of the first dimension reduction vector into a first neural network to obtain a first prediction deep plating capability grade vector;
obtaining a second loss function according to the first prediction deep plating capacity grade vector;
inputting the first N/2 dimension of the first dimensionality reduction vector into a first neural network to obtain a second prediction deep plating capacity grade vector;
acquiring a third loss function according to the second prediction deep plating capability grade vector;
and obtaining a first loss function according to the second loss function and the third loss function.
In the process of training the first neural network, a data set consisting of multi-dimensional vectors and deep plating capability grade vectors is adopted, neural network parameters are continuously updated in the training process, meanwhile, an initial dimensionality reduction matrix corresponding to the first dimensionality reduction vector is continuously updated and obtained until convergence, and the initial dimensionality reduction matrix when the first neural network converges is obtained and serves as a final dimensionality reduction matrix.
And separating the noise information in each multi-dimensional vector in the process of obtaining the second dimension-reducing vector according to the final dimension-reducing matrix, wherein the front N/2 dimension of each obtained second dimension-reducing vector contains the noise information of the corresponding multi-dimensional vector, and the rear N/2 dimension does not contain the noise information of the corresponding multi-dimensional vector.
In this embodiment, a randomly initialized H × N initial dimension reduction matrix is presetLet us orderThen, it isIs thatThe dimension of the first dimension-reduced vector is N dimension; i.e. H-dimensionalVector dimension reduction to N dimensionA first dimension reduction vector; constructing an FC first neural network by taking the first dimension reduction vector as an input and taking the deep plating capacity grade vector as an output; the embodiment determines an accurate final dimensionality reduction matrix by the following methodAnd letThe first dimension N/2 contains noise, and the second dimension N/2 does not contain noise, and the specific method is as follows:
first, willThe last N/2 dimension is input into a first neural network, and a ten-dimensional first prediction covering power grade vector is output(ii) a Due to the desire of the present embodimentThe last N/2 dimension of the first neural network does not contain data noise, and a first dimensionality reduction vector is input for the first neural network in consideration of the fact that the data without the noise can enable the first neural network to obtain an accurate deep plating capability levelIn this case of the last N/2 dimension, the second loss function is constructed as follows:
indicates that the plating depth of the kth electroplating process is graded asThe throwing power level vector of (2);
indicates that the plating depth of the kth electroplating process is graded asA first predicted throwing power level vector of; wherein, the first and the second end of the pipe are connected with each other,the deep plating capability grade corresponding to the kth electroplating process is represented and taken;
In the subsequent stepWhen the minimum value is taken, the output of the first neural network can be enabled to be ANDEqual, the deep plating capability grade of the network can be ensured, and further the purpose is achievedThe last N/2 dimension of (1) does not contain data noise.
Secondly, willThe first N/2 dimension is input into the first neural network, and a ten-dimensional second prediction deep plating capability grade vector is output. Due to the desire of the present embodimentThe first N/2 dimension of (2) contains data noise, and a first dimensionality reduction vector is input for the first neural network in consideration of the fact that the first neural network cannot obtain an accurate deep plating capability level after the data introduces noiseFor the first N/2-dimension case, the third loss function is constructed as follows:
indicates that the plating depth degree corresponding to the kth electroplating process isA second predicted throwing power level vector of (1); wherein, the first and the second end of the pipe are connected with each other,the deep plating capability grade corresponding to the kth electroplating process is represented and taken;
In the subsequent stepThe minimum value is taken to ensure that the result output by the first neural network has larger entropy, so that the first neural network cannot be operated according to the entropyThe front N/2 dimension of the steel plate obtains the deep plating capability grade, thereby achieving the purpose ofThe first N/2 dimensions of (1) contain the purpose of data noise.
Thirdly, acquiring a first loss function according to the second loss function and the third loss function; the specific first loss function calculation formula is as follows:
in the formula (I), the compound is shown in the specification,representing a first loss function;representing a second loss function;representing a third loss function;
Then, according to a gradient descent method, training a first neural network by using the loss function and a second data set, continuously updating parameters of the neural network in the training process, and simultaneously continuously updating a matrix D, and when the neural network converges, obtaining the value of a final dimensionality reduction matrix D when the first neural network converges;
it should be noted that, the specific process of constructing the first neural network is as follows:
inputting PCB electroplating process control parameters corresponding to the multidimensional vectors by the network, and outputting the deep plating capability grade vectors by the network;
the data used by the network training is collected PCB electroplating historical data in the production process, and the PCB electroplating process control parameters comprise: concentration of liquid medicine, current density, composition of light agent, vibration information (vibration frequency and amplitude) and jet flow information (jet flow speed);
and the network label is a deep plating capability grade vector corresponding to the deep plating capability corresponding to each electroplating process control parameter in the collected historical processing process of the PCB.
At this time obtainThe dimensionality reduction result ofWherein K =1,2, \8230;, K, this dimension reduction resultMedium to medium N/2 dimensional inclusionOf a noise ofIn the middle and later N/2 dimensions do not containOr, the embodiment enables the noise of (2) to be reduced by the first neural networkIn the storage of noise informationInformation in the medium-to-medium N/2-dimension without noise is stored inIn middle and later N/2 dimension, realizeAnd (4) separating the noise information. And the training of the second neural network with subsequent deep plating capability is facilitated.
S4, reducing the dimension of each multi-dimensional vector according to the final dimension reduction matrix to obtain a plurality of second dimension reduction vectors;
adopting the final dimension reduction matrix D and obtaining the vector corresponding to each electroplating process control parameterMultiplying to obtain the reduced frontDimension vector data, denoted asThe larger the data value of which dimension in the vector is, the more information noise the data on the dimension reflects.
To find out beforeThe information module in the information before dimensionality reduction corresponding to the dimensionality vector data calculates the abnormal degree of the information module before dimensionality reduction corresponding to the dimensionality according to each dimensionality data value of the noise vector after dimensionality reduction,
s5, acquiring the abnormal degree of any numerical value in the multi-dimensional vector corresponding to each second dimension-reducing vector according to the numerical value in the final dimension-reducing matrix and the front N/2-dimensional numerical value of each second dimension-reducing vector; the method comprises the following specific steps:
the abnormal degree of each electroplating process control parameter value is as follows:
in the formula (I), the compound is shown in the specification,the value of the jth row and jth column of the final dimension reduction matrix is represented, and the larger the value is, the more information of the jth dimension data value of the extracted original electroplating parameter data is;
the former N/2 dimensional data of the dimensionality reduction vector corresponding to the kth electroplating process control parameter is shown, wherein,;
representing the abnormal degree of the jth numerical value in the kth electroplating process control parameter;
and sequentially obtaining the abnormal degree of each data value in each process control parameter in the mode.
S6, constructing a second neural network according to the multidimensional vector as an input and the deep plating capability as an output;
acquiring a regularization coefficient weight value of the input layer neuron parameters in the second neural network according to the abnormal degree of any numerical value in each multi-dimensional vector; acquiring a comprehensive loss function according to the regularization coefficient weight value and the neuron parameter value of the input layer in the second neural network; training a second neural network based on a data set consisting of a comprehensive loss function, a multi-dimensional vector and a deep plating capability to obtain a trained second neural network;
it should be noted that, in order to prevent the characteristic of the abnormal data fitted by the network, that is, the second neural network is not accurate enough due to overfitting, the regularization coefficient weight of each network parameter needs to be adjusted according to the abnormal degree of each data, and the regularization coefficient weight of the neuron parameter corresponding to the data with a large degree is increased, that is, the neuron parameter corresponding to the data is as small as possible or even 0; fitting a deep plating capability prediction second neural network: and obtaining the abnormal degree of each electroplating parameter data through the network, so that the second neural network needs to be predicted again by combining the abnormal degree of each data, and the network is utilized to obtain the optimal electroplating viscosity control parameter. The network is a DNN network, the network structure is an FC regression network, the input of the network is the first data set obtained in the first step, the output of the network is a deep plating capacity value, and the traditional loss function is a mean square error loss function.
The method comprises the steps that a regularization loss function is obtained according to a regularization coefficient weight value and a neuron parameter value of an input layer in a second neural network in the process of obtaining the comprehensive loss function; and then obtaining the normalized loss function and the original mean square error loss function of the second neural network.
Specifically, in the process of obtaining the PCB plating process control parameters when the deep plating capability of the PCB is the maximum, on the basis of an objective function, when the parameters in the second neural network are kept unchanged, the magnitude of the numerical values in the multidimensional vector is continuously updated by adopting a gradient descent method to train the second neural network again until convergence, and the multidimensional vector in the final convergence is obtained, namely the PCB plating process control parameters when the multidimensional vector corresponds to the PCB with the maximum deep plating capability are obtained.
In this embodiment, because the collected data of the deep plating capability is not accurate enough, in order to prevent the overfitting phenomenon of the fitted second neural network of the deep plating capability, that is, the characteristics of the data with inaccurate deep plating capability are also fitted to cause the inaccuracy of the second neural network of the deep plating capability, the regularization coefficient needs to be controlled by combining the abnormal degree of the deep plating data, the regularization weight coefficient of the neural network parameter corresponding to the data with large abnormal degree is as large as possible, and further the neural network parameter value corresponding to the data with large abnormal degree is as small as possible, and the regularization coefficient weight calculation method of the neural network neural parameter of the second neural network of the input layer corresponding to each specific data is as follows:
in the formula (I), the compound is shown in the specification,representing the abnormal degree of the jth numerical value of the kth electroplating process control parameter data;
representing the regularization coefficient weight value of the neuron parameter corresponding to the jth dimension value of the kth electroplating process control parameter data;
the determination method for solving the weight of the regularization coefficient only adapts to the neuron parameters of the first layer;representing the normalized coefficient term. The regularization coefficient weights for the second and deeper layer neuron parameters are:
in the formula (I), the compound is shown in the specification,representing the abnormal degree of the jth numerical value of the kth electroplating process control parameter data;the average value of the abnormal degree of the jth numerical value in the k electroplating process control parameters is represented;
the regularization loss function is as follows:
in the formula (I), the compound is shown in the specification,the regularization coefficient weight of the jth neuron parameter corresponding to the kth electroplating process control parameter data is represented,representing the jth neuron parameter value.
Wherein, the aboveThe total number of the plating process control parameters is equal to H in S1, and j is the numerical ordinal number of the plating process control parameters.
The synthetic loss function is:
in the formula (I), the compound is shown in the specification,representing regularization loss functionCounting;representing the original mean square error loss function in the second neural network.
Training a second neural network based on a data set consisting of a comprehensive loss function, a multi-dimensional vector and a deep plating capability to obtain the trained second neural network;
it should be noted that, in the process of training the second neural network:
inputting PCB electroplating process control parameters corresponding to multidimensional vectors of the network, and outputting the network as deep plating capacity;
the data used by the network training is collected PCB electroplating historical data in the production process, and the PCB electroplating process control parameters comprise: concentration of liquid medicine, current density, composition of light agent, vibration information (vibration frequency and amplitude) and jet flow information (jet flow speed);
the label of the network is the deep plating capacity corresponding to each electroplating process control parameter in the collected historical processing process of the PCB.
S7, reducing the dimension of a multi-dimensional vector formed by any PCB electroplating process control parameter through a final dimension reduction matrix to obtain a third dimension reduction vector, and obtaining a target function according to the last N/2 dimension of the third dimension reduction vector as the input of a second neural network; and continuously updating the magnitude of the numerical value in the multi-dimensional vector based on the objective function, training the second neural network again, and obtaining the electroplating process control parameters of the PCB when the deep plating capacity of the PCB is maximum.
Specifically, in the process of obtaining the PCB plating process control parameters when the deep plating capability of the PCB is the maximum, on the basis of an objective function, when the parameters in the second neural network are kept unchanged, the magnitude of the numerical values in the multidimensional vector is continuously updated by adopting a gradient descent method, the second neural network is trained again until convergence is reached, and the multidimensional vector when final convergence is obtained, namely the PCB plating process control parameters when the multidimensional vector corresponds to the PCB with the maximum deep plating capability are obtained.
In this embodiment, it is desired to obtain a PCB electroplating process control parameter x, so that the deep plating capability of the PCB obtained under the process control parameter x is the maximum, and the specific method is as follows:
randomly initializing a process control parameter x, calculatingWill beInputting the obtained output result into the trained second neural network in the last N/2 dimension and recording the output result as y 0 。
This embodiment keeps the neural network parameters and the final dimensionality reduction matrix D fixed, then y 0 Is a mathematical model containing an unknown number x;
an objective function loss0= exp (-y) is then constructed 0 ) Thus, loss0 is also a mathematical model with an unknown x, the purpose of this example being to obtain a mathematical model of y 0 The maximum x is that an x is obtained, so that the loss0 is minimum, therefore, the gradient descent method is adopted to solve the x corresponding to the minimum value of the loss0, and the obtained x is the corresponding process control parameter when the deep plating capacity of the PCB is maximum.
It should be noted that, in the process of obtaining x, based on the objective function, while the parameters in the second neural network are kept unchanged, the magnitude of the numerical value in the x multidimensional vector is continuously updated by adopting a gradient descent method to train the second neural network again until convergence, and the multidimensional vector x in final convergence is obtained, that is, the PCB board electroplating process control parameters when the multidimensional vector x corresponds to the maximum deep plating capability of the PCB board are obtained.
The electroplating control is recommended to be carried out by using the process control parameter x in the subsequent PCB electroplating process, but the electroplating process has certain complexity, so that the fine adjustment can be carried out on the process control parameter according to the requirement in the actual electroplating process.
The present invention is not limited to the above preferred embodiments, and any modifications, equivalent substitutions, improvements, etc. within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims (9)
1. A method for improving the deep plating capability of a PCB is characterized by comprising the following steps:
acquiring control parameters of the electroplating process of the PCB for multiple times, and acquiring the deep plating capacity of the PCB in each electroplating process; the control parameters of the PCB electroplating process each time form multidimensional vectors;
discretizing according to the deep plating capability of the PCB in each electroplating process to obtain a deep plating capability grade vector;
reducing the dimension of the multi-dimensional vector according to a preset initial dimension reduction matrix to obtain a first dimension reduction vector; wherein the dimensionality of the first dimensionality reduction vector is N dimensionality;
constructing a first neural network according to the first dimension reduction vector as an input and the deep plating capability grade vector as an output; respectively taking the front N/2 dimension and the rear N/2 dimension in the first dimension reduction vector as the input of a first neural network to obtain a first loss function; training a first neural network based on a first loss function, and acquiring a final dimensionality reduction matrix when the first neural network is converged;
reducing the dimension of each multi-dimensional vector according to the final dimension reduction matrix to obtain a plurality of second dimension reduction vectors;
in the process of obtaining the second dimension-reducing vectors according to the final dimension-reducing matrix, the noise information in each multi-dimensional vector is separated, so that the front N/2 dimension of each obtained second dimension-reducing vector contains the noise information of the corresponding multi-dimensional vector, and the rear N/2 dimension does not contain the noise information of the corresponding multi-dimensional vector;
acquiring the abnormal degree of any numerical value in the multi-dimensional vectors corresponding to each second dimension-reducing vector according to the numerical value in the final dimension-reducing matrix and the front N/2-dimensional numerical value of each second dimension-reducing vector;
constructing a second neural network according to the multidimensional vector as input and the deep plating capability as output;
acquiring a regularization coefficient weight value of input layer neuron parameters in a second neural network according to the abnormal degree of any numerical value in each multi-dimensional vector; acquiring a comprehensive loss function according to the regularization coefficient weight value and the neuron parameter value of the input layer in the second neural network; training a second neural network based on a data set consisting of a comprehensive loss function, a multi-dimensional vector and a deep plating capability to obtain the trained second neural network;
reducing the dimension of a multi-dimensional vector formed by electroplating process control parameters of any PCB through a final dimension reduction matrix to obtain a third dimension reduction vector, and obtaining a target function by taking the last N/2 dimension of the third dimension reduction vector as the input of a second neural network; and continuously updating the magnitude of the numerical value in the multi-dimensional vector based on the objective function, and training the second neural network again to obtain the PCB electroplating process control parameter when the deep plating capacity of the PCB is maximum.
2. The method for improving the throwing power of the PCB according to claim 1, wherein in the training of the first neural network, a data set consisting of a multidimensional vector and a throwing power grade vector is adopted, the neural network parameters are continuously updated in the training process, an initial dimensionality reduction matrix corresponding to the first dimensionality reduction vector is continuously updated until convergence, and the initial dimensionality reduction matrix when the first neural network converges is obtained as a final dimensionality reduction matrix.
3. The method for improving the deep plating capability of the PCB of claim 1, wherein in the process of obtaining the comprehensive loss function, the regularization loss function is obtained according to the weight value of the regularization coefficient and the neuron parameter values of the input layer in the second neural network; and obtaining the normalized loss function and the original mean square error loss function of the second neural network.
4. The method for improving the deep plating capability of the PCB according to claim 1, wherein in the process of obtaining the control parameters of the PCB plating process when the deep plating capability of the PCB is maximized, based on an objective function, when the parameters in the second neural network are kept unchanged, the magnitude of the numerical values in the multidimensional vector is continuously updated by adopting a gradient descent method to train the second neural network again until convergence, and the multidimensional vector when final convergence is obtained, namely the control parameters of the PCB plating process when the multidimensional vector corresponds to and maximizes the deep plating capability of the PCB are obtained.
5. The method for improving the throwing power of PCB board according to claim 1, wherein the first loss function is obtained according to the following steps:
inputting the last N/2 dimension of the first dimension reduction vector into a first neural network to obtain a first prediction deep plating capability grade vector;
obtaining a second loss function according to the first prediction deep plating capacity grade vector;
inputting the first N/2 dimension of the first dimension reduction vector into a first neural network to obtain a second prediction deep plating capability grade vector;
obtaining a third loss function according to the second prediction deep plating capacity grade vector;
and obtaining a first loss function according to the second loss function and the third loss function.
6. The method for PCB throwing power improvement of claim 5, wherein the second loss function is as follows:
indicates that the plating depth degree corresponding to the kth electroplating process isThe throwing power level vector of (2);
7. The method for PCB board throwing power improvement of claim 5, wherein the third loss function is as follows:
8. The method for improving the throwing power of the PCB according to claim 1, wherein the throwing power of the PCB in each electroplating process is obtained according to the following steps:
the plating thickness of PCB diaphragm orifice inner wall and the epitaxial plating thickness in PCB diaphragm orifice are detected through cladding material thickness detector, obtain the deep-plating ability according to the ratio of the plating thickness in diaphragm orifice inner wall and the epitaxial plating thickness in diaphragm orifice.
9. The method for improving the throwing power of the PCB board according to claim 1, wherein the throwing power grade vector is obtained according to the following steps:
discretizing the deep plating capability of the PCB in each electroplating process into a plurality of deep plating capability quality grades;
and acquiring the quality grade of the deep plating capability corresponding to the deep plating capability of the PCB in each electroplating process, and performing one-hot coding on the quality grade of the deep plating capability to acquire a deep plating capability grade vector corresponding to each quality grade of the deep plating capability.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202211177716.XA CN115261963A (en) | 2022-09-27 | 2022-09-27 | Method for improving deep plating capability of PCB |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202211177716.XA CN115261963A (en) | 2022-09-27 | 2022-09-27 | Method for improving deep plating capability of PCB |
Publications (1)
Publication Number | Publication Date |
---|---|
CN115261963A true CN115261963A (en) | 2022-11-01 |
Family
ID=83757687
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202211177716.XA Pending CN115261963A (en) | 2022-09-27 | 2022-09-27 | Method for improving deep plating capability of PCB |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN115261963A (en) |
Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101539781A (en) * | 2009-04-22 | 2009-09-23 | 北京中冶设备研究设计总院有限公司 | Electrogalvanizing zinc coating thickness BP neural network control method and application in PLC thereof |
CN110503198A (en) * | 2019-07-23 | 2019-11-26 | 平安科技(深圳)有限公司 | Obtain method, apparatus, equipment and the storage medium of neural network test report |
CN111261146A (en) * | 2020-01-16 | 2020-06-09 | 腾讯科技(深圳)有限公司 | Speech recognition and model training method, device and computer readable storage medium |
CN111860495A (en) * | 2020-06-19 | 2020-10-30 | 上海交通大学 | Hierarchical network structure searching method and device and readable storage medium |
CN114705747A (en) * | 2022-03-31 | 2022-07-05 | 生益电子股份有限公司 | Method for monitoring deep plating capability based on volt-ampere cycle method |
CN115019049A (en) * | 2022-08-08 | 2022-09-06 | 四川大学 | Bone imaging bone lesion segmentation method, system and equipment based on deep neural network |
CN115066019A (en) * | 2022-06-10 | 2022-09-16 | 东南大学 | Intelligent beam selection performance improving method based on gradient mask |
-
2022
- 2022-09-27 CN CN202211177716.XA patent/CN115261963A/en active Pending
Patent Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101539781A (en) * | 2009-04-22 | 2009-09-23 | 北京中冶设备研究设计总院有限公司 | Electrogalvanizing zinc coating thickness BP neural network control method and application in PLC thereof |
CN110503198A (en) * | 2019-07-23 | 2019-11-26 | 平安科技(深圳)有限公司 | Obtain method, apparatus, equipment and the storage medium of neural network test report |
CN111261146A (en) * | 2020-01-16 | 2020-06-09 | 腾讯科技(深圳)有限公司 | Speech recognition and model training method, device and computer readable storage medium |
CN111860495A (en) * | 2020-06-19 | 2020-10-30 | 上海交通大学 | Hierarchical network structure searching method and device and readable storage medium |
CN114705747A (en) * | 2022-03-31 | 2022-07-05 | 生益电子股份有限公司 | Method for monitoring deep plating capability based on volt-ampere cycle method |
CN115066019A (en) * | 2022-06-10 | 2022-09-16 | 东南大学 | Intelligent beam selection performance improving method based on gradient mask |
CN115019049A (en) * | 2022-08-08 | 2022-09-06 | 四川大学 | Bone imaging bone lesion segmentation method, system and equipment based on deep neural network |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110263846B (en) | Fault diagnosis method based on fault data deep mining and learning | |
CN106407464A (en) | KNN-based improved missing data filling algorithm | |
CN111144552B (en) | Multi-index grain quality prediction method and device | |
CN110096810B (en) | Industrial process soft measurement method based on layer-by-layer data expansion deep learning | |
CN110889085A (en) | Intelligent wastewater monitoring method and system based on complex network multiple online regression | |
CN110390561B (en) | User-financial product selection tendency high-speed prediction method and device based on momentum acceleration random gradient decline | |
CN113570000A (en) | Ocean single-factor observation quality control method based on multi-model fusion | |
CN108510180B (en) | Method for calculating performance interval of production equipment | |
CN110379521B (en) | Medical data set feature selection method based on information theory | |
CN105786711A (en) | Data analysis method and device | |
CN111931983A (en) | Precipitation prediction method and system | |
CN112085252A (en) | Counterfactual prediction method about set type decision effect | |
CN110110447B (en) | Method for predicting thickness of strip steel of mixed frog leaping feedback extreme learning machine | |
CN113627594B (en) | One-dimensional time sequence data augmentation method based on WGAN | |
CN115261963A (en) | Method for improving deep plating capability of PCB | |
Ivanovski et al. | Fruit firmness prediction using multiple linear regression | |
CN116029604B (en) | Cage-raised meat duck breeding environment regulation and control method based on health comfort level | |
CN117000780A (en) | Self-adaptive cold-rolled strip multichannel shape forecasting method | |
CN106599391B (en) | Association vector machine soft measurement modeling method based on dynamic weighting of triangle angle values | |
Ergashev | Should risk managers rely on maximum likelihood estimation method while quantifying operational risk? | |
CN115738285A (en) | Game quality evaluation feedback method and system | |
CN108053266A (en) | A kind of patent value predictor method and device | |
CN114021644A (en) | Method for calculating regional representative underground water level based on K-means and de-clustering method | |
CN113035363B (en) | Probability density weighted genetic metabolic disease screening data mixed sampling method | |
CN114384870B (en) | Complex industrial process running state evaluation method based on nuclear local linear embedded PLS |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination |