TWI676939B - Electronic component packaging classification system using neural network for classification - Google Patents

Abstract

An electronic component packaging classification system using neural network-like classification for classification. The electronic component packaging system includes a service database, an external database, a feature selection module, a data integration module, and a classification processing module. The service database is used for external input of electronic component drawings. The external database stores electronic component package type information. The feature selection module records the electronic component package type characteristics. The data integration module performs data pre-processing on the selected feature values. Processing and normalization to obtain the data to be classified, the classification processing module receives the data to be classified and displays the classification result in the service database.

Description

Electronic Zeros Using Neural Networks for Classification Package classification system

The invention relates to a classification system, in particular to an electronic component packaging classification system using a neural network to classify.

With the development of technology today, the electronic circuit design and assembly workflow has gradually become automated. In the process of designing printed circuit boards, you must first import the Footprint Library and set the PCB Parameters Setup. ), Placement (routing), routing (Routing) and finally enter the manufacturability design check (Design for Manufacture Check, DFM Check) and other stages.

Before performing the manufacturability design check, the traditional method is to manually sort the package types of the electronic parts used on the printed circuit board one by one manually by the layout engineer, and the layout engineers judge the packaging type of the electronic parts mostly based on electronics. The name of the part drawing and the appearance are judged by the number of pins and the placement of the pins. This process depends not only on the engineer's own work experience, but also cannot ensure the correctness of the electronic component package type classification.

With the evolution of packaging technology, the packaging types of electronic parts are becoming more and more variable, and the patterns of some packaging types are very similar. For the layout engineer, it is more difficult to identify the package type through the electronic part pattern, and if there is an error in the judgment of the electronic part package type, it will affect the layout engineer's workflow and even the production yield and product quality of the assembly plant.

The above shortcomings all show various problems arising from the operation of conventional electronic component packaging classification. Therefore, it is inevitable to develop packaging classification tools to assist layout engineers in reducing the probability of electronic component packaging type classification errors.

In view of this, an object of the present invention is to provide an electronic component package classification system using neural network-like classification. The electronic component package classification system includes a service database, an external database, a feature selection module, A data integration module and a classification processing module.

The service database is used for external input of electronic part patterns, and training data for receiving and storing associated input and output data. The external database stores a plurality of electronic part package type data, the feature selection module and the External database connection with electronic zero recorded The package type characteristics of the package are based on the electronic part drawings to be classified according to the service database input. The feature selection module performs feature selection from the external database according to the package type characteristics.

The data integration module performs data preprocessing and normalization on the feature values selected by the feature selection module to remove error noise from the data and fill in missing data, and limit the feature value distribution of the feature to a specific interval To obtain the data to be classified, the classification processing module receives the data to be classified and displays the classification result in the service database.

Another technical means of the present invention is that the above-mentioned classification processing module includes a processor storing a command for executing an action, and the action includes: the user inputting an electronic part drawing to be classified into the service database; The feature selection module performs feature selection from the external database according to the package type characteristics of the electronic part pattern; the data integration module performs data preprocessing and normalization on the selected feature values to obtain data to be classified; and the The service database obtains the classification results of the package types of electronic parts.

Another technical means of the present invention is that the above-mentioned electronic component package classification system further includes a training module and a parameter storage module, and the training module is connected to the data integration module and the service database, The training scale and neural network parameters of the training data set for training are determined as the basis for subsequent classification. Among them, the convergence condition of training is that the cumulative error after the current training is less than a given threshold, that is, the training is stopped, and the The parameter storage module is connected to the training module and the service database to record training parameter data used by the training module.

,v _a <v _b 關係式，其中，v'為正規化至v _a 、v _b 後的特徵值，v為需作正規化之特徵值，v _max 為一項特徵中的最大特徵值，而v _min 為一項特徵中的最小特徵值。 Another technical means of the present invention is that the above-mentioned data integration module normalizes the feature values into the intervals of v _a and v _b , and satisfies v ' =

, V _a <v _b relation, where, v 'is normalized to the feature value of v _a, v _b, v is the need for the normalized feature values, v _max is the maximum eigenvalue of a feature, and v _min is the minimum eigenvalue of a feature.

Another technical means of the present invention is that the above-mentioned neural network parameters are convergence conditions, the number of hidden layer neurons, the number of hidden layers, the initial learning rate, the initial momentum, the threshold value, the weight value, and the partial weight value. Either or a combination.

Another technical means of the present invention is that the number of neurons in the hidden layer j above meets ( x × ( input + output )), 1.5 < x <2, where input is 19 features of the input package type and output is classification The output package type is 10.

Another technical means of the present invention is the above-mentioned seal The package type data records any one or a combination of the above-mentioned electronic component appearance information, printed circuit board restricted area information, solder joint information, geometric shape parameters, applicable field parameters, electrical parameters, and contact parameters.

Another technical means of the present invention is that the above-mentioned package type features include the physical appearance of the electronic part, the physical pins of the electronic part, and the electronic part pattern.

Another technical means of the present invention is that the weighting ratio of the aforementioned package type features is that the electronic part pattern is larger than the physical appearance of the electronic part, and the physical appearance of the electronic part is greater than the physical pin of the electronic part.

Another technical means of the present invention is that the physical appearance of the electronic component, the physical pins of the electronic component, and the electronic component pattern are selected from the number of pins of the electronic component, the original physical length of the electronic component, and the maximum physical length of the electronic component. 、 Minimum electronic component entity length, original electronic component entity width, maximum electronic component entity width, minimum electronic component entity width, electronic component entity height, electronic component entity and circuit board spacing, large electronic component pin length , Small electronic part pin length, large electronic part pin width, small electronic part pin width, large electronic part pattern pin length, small electronic part pattern pin length, large electronic part pattern pin Width, small electronic part pattern pins There are 19 types including width, X-axis direction of pin pitch of electronic parts pattern, and Y-axis direction of pin pitch of electronic parts pattern.

The beneficial effect of the present invention is that the neural network is trained through the characteristics of the electronic part entity to find the training scale and neural network parameters that are most suitable for the classification system, and the accuracy of the normalized training result is lower than that of the normalized one. The training result is high, and it can solve the problems of easy judgment, time-consuming, and classification process that rely on the layout engineer's own work experience to manually classify the electronic component package types. It can also obtain higher quality training and classification results.

1‧‧‧Service Database

3‧‧‧ External Database

4‧‧‧Feature Selection Module

5‧‧‧Data Integration Module

6‧‧‧ Training Module

7‧‧‧parameter storage module

8‧‧‧Classification processing module

91 ~ 94‧‧‧step

FIG. 1 is a block diagram illustrating a preferred embodiment of the electronic component packaging classification system using neural network-like classification according to the present invention. FIG. 2 is a schematic diagram illustrating the architecture of the preferred embodiment at a node output calculation stage. Figure 3 is a schematic diagram illustrating the training process of the preferred embodiment; Figure 4 is a schematic diagram illustrating the architecture of the preferred embodiment in a weight correction phase; and Figure 5 is a schematic diagram illustrating the preferred embodiment A classification processing model in the embodiment The group executes a flow of action instructions.

The features and technical contents of the related patent application of the present invention will be clearly presented in the following detailed description of the preferred embodiments with reference to the drawings.

Referring to FIG. 1, a preferred embodiment of an electronic component packaging classification system using neural network-like classification according to the present invention. The electronic component packaging classification system includes a service database 1, an external database 3, and a feature selection module. 4. A data integration module 5, a training module 6, a parameter storage module 7, and a classification processing module 8.

The service database 1 is used for externally inputting electronic part patterns to be trained or classified, and to receive and store training data of associated input and output data, wherein the electronic part patterns are implemented by Electronic Design Automatic (Electronic Design Automatic, EDA) tool conversion file format.

The external database 3 stores package type data of a plurality of electronic parts, wherein the package type data records electronic part appearance information, printed circuit board restricted area information, solder joint information, geometric shape parameters, applicable field parameters, electrical Parameters, contact parameters, etc. Its combination.

The feature selection module 4 is connected to the external database 3, and records the package type characteristics of the electronic parts. According to the electronic part drawings to be trained or classified according to the service database 1, the feature selection module 4 is based on the package. Type features are selected from the external database 3.

The joining of electronic parts and circuit boards can roughly divide the packaging technology into Through Hole Technology (THT) and Surface Mount Technology (SMT). Here, the basic SMT type of electronic component packaging is based on the pin form. , Pin type, size, and functionality are divided into 44 categories, and this embodiment takes the common 25 categories and classifies them into 10 package types to meet the needs of layout engineers in determining package types.

At this stage, the feature selection module 4 obtains 19 features from the 25 types of SMT package types, and obtains the feature values, and prepares these feature values for data preprocessing by the data integration module 5.

Further, the package type features include a total of 19 types including the physical appearance of electronic parts, the physical pins of electronic parts, and the drawings of electronic parts. Among them, the physical appearance characteristics of the electronic part are the number of pins of the electronic part, Original electronic part physical length, maximum electronic part physical length, minimum electronic part physical length, original electronic part physical width, maximum electronic part physical width, minimum electronic part physical width, electronic part physical height, and electronic part The distance between the entity and the circuit board.

Electronic part physical pins are characterized by a large electronic part pin length, a small electronic part pin length, a large electronic part pin width, and a small electronic part pin width. The electronic part pattern features are the X-axis of the large electronic part pattern pin length, the small electronic part pattern pin length, the large electronic part pattern pin width, the small electronic part pattern pin width, and the electronic part pattern pin pitch. Direction, and the Y-axis direction of the pin pitch of the electronic part pattern.

Here, the weighting ratio of the package type feature is that the electronic part pattern is larger than the physical appearance of the electronic part, and the physical appearance of the electronic part is larger than the physical pin of the electronic part.

The data integration module 5 performs data preprocessing and normalization on the feature values selected by the feature selection module 4 to remove error noise from the data and fill in missing data, and limit the feature value distribution of the feature to one. Specific interval to get the training data set. Here, when the data processed by the data integration module 5 is a pattern of electronic parts to be trained, it is called training The training data set is used for training by the training module 6. If the data processed by the data integration module 5 is an electronic part pattern to be classified, it is called to-be-classified data as the classification result of the package type.

Preprocessing is to integrate data, clean up and fill in missing data, and convert data. Among them, the purpose of data integration is to solve the problems of inconsistent data, different units and duplicate data caused by training data composed of multiple different databases. If the data is inconsistent, it may be difficult to converge or affect the training result due to the different representation of the field content, so that the data forms a data set that is not conducive to training. Therefore, data integration is the first step in data preprocessing.

Furthermore, the purpose of data cleaning and filling of missing data is to confirm the completeness, correctness and rationality of the data. Due to the multiple sources of data, it is necessary to check whether the feature values are reasonable at this stage. The features selected here are the parameters of the electronic components , So use the overall average to fill in the missing data.

The purpose of data conversion is to convert the content of the data into easy-to-train or enhance the credibility of the training results. The work at this stage includes data generalization, the establishment of new attributes and data normalization. Data generalization refers to the improvement of the concepts represented by data. And meaning to reduce the type of feature value contained in the feature, and creating a new attribute means using the old attribute to find the training place New attributes required. Data normalization refers to converting data recorded under different standards or units into the same standard. After normalization, the data will be redistributed in a specific and smaller interval in order to improve the accuracy of the training results. Common normalization methods include extreme value normalization, Z-score normalization, and decimal normalization.

,v _a <v _b 關係式，其中，v'為正規化至v _a 、v _b 後的特徵值，v為需作正規化之特徵值，v _max 為一項特徵中的最大特徵值，而v _min 為一項特徵中的最小特徵值。 Here, the data integration module 5 normalizes the eigenvalues into the intervals of v _a and v _b , satisfying

, V _a <v _b relation, where, v 'is normalized to the feature value of v _a, v _b, v is the need for the normalized feature values, v _max is the maximum eigenvalue of a feature, and v _min is the minimum eigenvalue of a feature.

This embodiment uses a normalized and unnormalized training data set as an experimental comparison. The number of features, data amount, and number of output nodes selected in this data set and the training conditions of a neural network-like (also known as neural network) As shown in Table 1 below.

See Table 2 below for the training results of the normalized training data set. Table 2 below shows the training results for the non-normalized training data set. Here i - j - k is used to describe the neural network architecture, where i represents the input layer neural network. Number of neurons; j represents the number of neurons in the hidden layer; k represents the number of neurons in the output layer.

From the results in Tables 2 and 3 above, it can be seen that the average correct rate of the normalized data set No (19-50-10) is 99.2%, and the average correct rate of the unnormalized data set No (19-53-10) It is 51.8%. Therefore, the classification result of the normalized data set performs better, and it is 55.9% higher than that of the unnormalized data set.

In addition, the feature value distance of each feature is shortened after the data set is normalized, so that the neural network can more easily calculate the weight value of the connection between neurons through the normalized feature value during training. If the data is not normalized, the weight value may not be adjusted correctly because the weight value exceeds the interval of the activation function, causing the neural network to converge prematurely without achieving the effects of training and learning.

The present invention uses extreme value normalization to redistribute feature values in specific intervals to improve the efficiency of training neural networks, and the accuracy rate of training results after normalization is higher than that of non-normalized training results.

The training module 6 uses a feed-forward neural network (FNN) architecture with a backward transfer algorithm. The backward transfer algorithm belongs to a multilayer feedforward network and divides the neural network into an input layer (Input Layer), Hidden Layer and Output Layer. The input layer is used as an end to receive data and input messages in the network structure. How many neurons in the input layer represent how many different training features are used to represent the variables of the network input.

The hidden layer is located between the input layer and the output layer, and is used to show the interaction between the units. The number of hidden neurons is best found by trial and error. When the number of hidden neurons is larger, the convergence speed is slower and the error value is smaller. The output layer is used in the network architecture to process training results and output information, which is used to represent the variables output by the network.

，

，

...

表示輸入層神經元與隱藏層神經元相連的權重值；

則是隱藏層神經元之偏權值；h ₁，h ₂， h ₃...h _j 則是輸入項x _n 與權重值

之乘積總和，如公式

The inverse transfer algorithm is to minimize the error value and find the relationship between the connection weights between the input layer, the hidden layer and the output layer. As shown in Figure 2, the inverse transfer neural architecture can be divided into Input and Weight values Weight) and Activation Function. The weight value can be divided into weight value and partial weight value. Among them, x ₁ , x ₂ , x ₃ ... x _i are input signals;

,

,

...

Represents the weight value of the input layer neuron and the hidden layer neuron;

Is the partial weight of the hidden layer neurons; h ₁ , h ₂ , h ₃ ... h _j are the input terms x _n and the weight values

Product sum, such as formula

Then, h _{j is} substituted into the activation function f ^{tanh to} generate the output of the hidden layer and simultaneously used as the input of the next layer. In order to simulate the operation mode of biological neural networks, the activation function is usually a non-linear transformation. The traditional activation functions are Hyper Tangent function and Sigmoid function [28], as follows:

，

，

...

表示隱藏層神經元與輸出層神經元相連的權重值；

則是輸出層神經元之偏權值；O ₁,O ₂,...,O _k 則是輸入項f ^tanh (h _j )與權重值

之乘積總和，最後將O _k 代入活化函數f ^sig (O _k )，產生神經元輸出y _k ，如公式y _k ( O )=f ^sig (O _k )。 The activation function used by the hidden layer is the Hyper tangent function, and the output is the Sigmoid function.

,

,

...

Represents the weight value of the hidden layer neuron and the output layer neuron;

Are the partial weights of the output layer neurons; O ₁ , O ₂ , ..., O _k are the input terms f ^tanh ( h _j ) and the weight values

The sum of the product, and finally substituting O _k activation function f ^sig (O _k), generates an output neurons y _k, as shown in equation ^{_{y k (O) = f sig}} (O k).

When the backward transfer neural network does not reach the convergence condition, it will calculate the error between the output result and the target result and adjust the weights to retrain until the convergence condition is reached, such as the formula w _t = ( w _{t -1} + △ w ).

The training module 6 is connected to the data integration module 5 and the service database 1, and determines the training scale and neural network parameters for training on the training data set as a basis for subsequent classification, and sends the training results to the Service database 1, where the convergence condition of training is that the cumulative error after the current training is less than a given threshold, that is, training is stopped.

Referring to FIG. 3, in the preferred embodiment, the training process is divided into a network initialization phase, a node output calculation phase, and a weight correction phase. This node is also referred to as a neuron. First, through the network initialization phase, training data (herein referred to as the training data set), setting network input parameters, randomly generating weights (Weights) and partial weights (Bias), assigning weights and partial weights, and then , Ready to enter the node output calculation stage, this stage includes calculating the node output value of the hidden layer, applying the activation function of the hidden layer (Hyper tangent), calculating the node output value of the output layer, and applying the activation function of the output layer node (Sigmoid). Finally, the error correction gradient is calculated through the weight correction stage, and the weighting value and partial weight value, the learning rate is adjusted to the output result that reaches the convergence standard, and the training process is terminated. The training process ends after the weight correction phase and node output calculation phase reach the output result of the convergence standard.

Further, during the network initialization phase, the system It will first ask for neural network parameters and initialize the weights and partial weights. Three neural network parameters are set at this stage: initial learning rate, initial momentum, and number of hidden layer nodes.

Initial learning rate: When the initialization is performed, the learning rate is set to the interval of [0,1]. Here, the adaptive learning rate adjustment method is used, and the reference is made to the accumulated error value to determine whether the training direction is correct. If the error value shows a downward trend, it indicates that the training direction is correct, it will increase the learning rate and speed up the learning speed, otherwise it will add a penalty factor to reduce the learning rate and reduce the learning steps to modify the training direction.

Initial momentum: In addition to the setting of the learning rate, the size of the momentum also affects the learning efficiency of the neural network. The main function of momentum is to stabilize the shock caused by calculating the weight value after the learning rate is adjusted. During the initialization process, the parameters can be set in the interval [0,1] as the learning rate. The system will automatically add this parameter to adjust each time the learning rate and weight value are adjusted.

Number of nodes in the hidden layer: The number of nodes in the hidden layer will affect the convergence rate, learning efficiency, and training results. Here, the nodes in the hidden layer choose to use trial and error.

Convergence conditions can be set up to the maximum number of training or tired Stop counting when the error is less than the given threshold. Among them, the maximum number of trainings is used to stop when the training number reaches the set maximum number of trainings, which indicates that the training cannot make the neural network converge. You need to readjust the neural network. Or check the training data set for abnormality. As long as one of the above two conditions is met, the training process is terminated.

即停止訓練，

為當次訓練累積之誤差均方根RMSE，

為前次訓練累積之誤差均方根RMSE，滿足關係式：

，其中

與

滿足關係式：

，v ^rmse 為每次訓練結果後所累積之誤差均方根RMSE，c _d 為訓練資料集的資料量，c _o 為神經網路的輸出位元數，

為分類結果的目標值，

為當前分類結果的近似值。 Here, the convergence condition for training is that the cumulative error after the current training is less than the previous one

Stop training,

Is the root mean square error RMSE accumulated during the training,

The root mean square RMSE of the error accumulated from the previous training, satisfying the relationship:

,among them

versus

Satisfaction relationship:

^, V rmse results after training for each of the accumulated rms error RMSE, c _d is the amount of data the training data set, c _o is the number of bits output neural network,

Is the target value of the classification result,

Is the approximate value of the current classification result.

Further, the parameters of the neural network are any one or a combination of the above, such as convergence conditions, the number of hidden layer neurons, the number of hidden layers, the initial learning rate, the initial momentum, the threshold value, the weight value, and the partial weight value.

Please refer to Figure 2 again. During the node output calculation phase, the output value of each input node is calculated step by step, and added after the calculation is completed. After the partial weight value passes through the activation function, it is used as the input value of the next layer.

計算h _j ，使用公式h _j ( X )=

。接著將h _j 經過活化函數f ^tanh 後所得到的值作為隱藏層連結輸出層的輸入值，再乘上隱藏層連結輸出層的權重值

可得O _k ，使用公式O _k ( H )=

。最後在經由活化函數，得到每一筆資料的分類結果y _k ，使用公式y _k ( O )=f ^sig (O _k )。 Here i - j - k is used to describe the neural network architecture, where i represents the number of neurons in the input layer; j represents the number of neurons in the hidden layer; k represents the number of neurons in the output layer. x ₁ ~ x _i are the input feature values, and the weight values of the hidden layer are connected through the input layer

Calculate h _j using the formula h _j ( X ) =

. Then use the value of h _j after the activation function f ^tanh as the input value of the hidden layer link output layer, and then multiply the weight value of the hidden layer link output layer.

O _k can be obtained using the formula O _k ( H ) =

. Finally, through the activation function, the classification result y _{k of} each piece of data is obtained, using the formula y _k ( O ) = f ^sig ( O _k ).

With reference to FIG. 4, when the weight correction phase is entered, the weight value, partial weight value, and learning rate will be adjusted with reference to the cumulative error after the previous training. By adjusting these three variables, the training module 6 has better learning ability. In addition, the training conditions can be slightly modified through the results of each training to ensure that the learning direction is correct and the best learning effect can be obtained.

The method of adjusting weights is to gradually calculate from the output layer to the input layer, respectively calculating the output layer partial weight gradient, the hidden layer to output layer weight value gradient, the hidden layer partial weight value gradient, and the hidden layer to input layer weight value gradient. After a total of four gradients, calculate the amount of change by the gradient, and finally Correct the weight value through the amount of change and momentum.

、輸出層到隱藏層每一節點的梯度

、隱藏層偏權值梯度

以及隱藏層到輸入層每一節點的梯度

，

為第k個輸出的目標值；

則是第k個輸出的近似值，使用公式為

、

、

、

When adjusting the weight, you need to calculate the partial weight gradient of the output layer first

Gradient from output node to hidden node

Hidden layer partial weight gradient

And the gradient from each node of the hidden layer to the input layer

,

Is the target value of the k- th output;

Is the approximate value of the k- th output, using the formula:

,

,

,

、輸出層至隱藏層權重值之變動量

、隱藏層偏權值

變動量以及輸入層至隱藏層權重值變動量

。此計算過程中，會乘上學習率η使變動量調整幅度更加明顯，使用公式為

、

、

、

。 Then calculate the change in partial weight of the output layer

Change amount of weight value from output layer to hidden layer

Hidden layer partial weights

The amount of change and the amount of change from the input layer to the hidden layer weight value

. In this calculation process, the learning rate η will be multiplied to make the adjustment of the fluctuation amount more obvious. Using the formula is

,

,

,

.

、隱藏層偏權值

、隱藏層連結輸出層之權重值

，以及輸出層偏權值

並乘上動量M ^mom ，以減緩因調整權重值而造成訓練過程的震盪，作為下一次訓練時使用之參數，公式為

、

、

、

。 Finally, the weight value of the input layer and the hidden layer can be updated through the gradient and the change of the weight value.

Hidden layer partial weights

The weight value of the hidden layer connected to the output layer

, And the output layer partial weight

And multiply it by the momentum M ^mom to reduce the shock of the training process caused by adjusting the weight value. As a parameter used in the next training, the formula is

,

,

,

.

與當次訓練結果的

比較，判斷當次的學習方向是否正確。若學習方向正確，則會為學習率加上獎勵因子，使下一次的學習可以更加快速，使學習過程得以更早達成收斂條件，反之則加上懲罰因子，使學習速度減慢，以維持學習效果，公式為：

In this stage, the adaptive learning rate is used as a factor for calculating the variation of the weight value. The adjustment of the learning rate will refer to the previous training result.

With the results of the training

Compare and judge whether the current learning direction is correct. If the learning direction is correct, a reward factor will be added to the learning rate, so that the next learning can be faster and the learning process can reach the convergence condition earlier. Otherwise, a penalty factor is added to slow the learning speed to maintain learning Effect, the formula is:

The RMSE value obtained through the training results during each training process can be used to adjust the weight and learning rate, so that the training process advances in the correct direction, so that it cannot converge during the training process.

In the preferred embodiment, the training scale includes an input layer, a hidden layer, and an output layer, where the input layer is the number of input package type features, the number of hidden layers is 1, and the output layer is a classification The number of output package types. There are 19 input layer features, and the output type of the classification is 10.

The number of hidden layer neurons j satisfies ( x × ( input + output )), 1.5 < x <2, where input is 19 features of the input package type and output is 10 of the classification output type. Preferably, when the value of the hidden layer neuron is close to the above formula, better training and training classification results can be obtained.

The output package types are Ball Grid Array (BGA), Quad Flat Package (QFP), Quad Flat No-Lead (QFN), Small Outline Integrated Transistor (SOT), Small Outline Integrated Circuit (SOIC), Small Outline Integrated Circuit No-Lead (SON) , Dual Flat No-Lead (DFN) packages on both sides, Small Outline Diode (SOD) packages, Small chip components Chip, and Surface-attached color-coded resistors (Metal Electrode Leadless Face (MELF), etc.

The parameter storage module 7 is connected to the training module 6 and the service database 1 to record training parameter data used by the training module 6.

Referring to FIG. 5, the classification processing module 8 receives the to-be-classified Information and display the classification results in the service database 1. In actual implementation, the classification processing module 8 may be separately disposed on the user end or in the same electronic device for training and classification of electronic component packaging types, and should not be limited to this. The classification result can be the data to be classified or the processed result of the data to be classified.

The classification processing module 8 includes a processor storing an execution instruction, and the operation includes the following steps.

First, step 91 is performed, and the user terminal inputs the electronic part pattern to be classified into the service database 1. Then, step 92 is performed. The feature selection module 4 is obtained from the external database according to the package type characteristics of the electronic part pattern. 3 Perform feature selection.

Then, step 93 is performed. The data integration module 5 performs data preprocessing and normalization on the selected feature values to obtain data to be classified.

Finally, step 94 is performed, and the service database 1 obtains the classification result of the package type of the electronic component.

The invention uses the features of 19 electronic component entities to train a neural network to find the training scale and neural network parameters that are most suitable for the classification system. The normalized training result has a higher accuracy rate than the non-normalized training result. Furthermore, when the number of hidden layer neurons satisfies ( x × ( input + output )), 1.5 < x <2, better training and training classification results can be obtained.

To sum up, the present invention uses a neural network-like electronic component packaging classification system to classify the service database 1, the external database 3, the feature selection module 4, the data integration module 5, The training module 6, the parameter storage module 7, and the classification processing module are mutually set, and combined with a reverse transfer neural network, it solves the error, time-consuming and classification easily caused by the manual classification of electronic component packaging types. The problem that the process is too dependent on the layout engineer's own work experience can also obtain higher quality training and classification results, so it can indeed achieve the purpose of cost invention.

However, the above are only the preferred embodiments of the present invention. When the scope of implementation of the present invention cannot be limited by this, that is, the simple equivalent changes and modifications made according to the scope of the patent application and the description of the invention, All are still within the scope of the invention patent.

Claims (10)

An electronic part packaging classification system using neural network-like classification for classification. The electronic part packaging classification system includes: a service database for external input of electronic part patterns, and training for receiving and storing associated input and output data. Data; an external database that stores the package type data of multiple electronic parts; a feature selection module that is connected to the external database and records the package type characteristics of the electronic parts, and is classified according to the desires entered by the service database Electronic component drawings, the feature selection module performs feature selection from the external database according to the package type feature; a data integration module performs data preprocessing and normalization of the feature values selected by the feature selection module to Clear error data from the data and fill in missing data, and limit the eigenvalues of the feature to a specific interval to obtain the data to be classified; and a classification processing module that receives the data to be classified and displays the classification results in The service database. According to the electronic component package classification system according to item 1 of the scope of patent application, wherein the classification processing module includes a processor storing an execution instruction, the action includes: the user inputting the electronic component to be classified Drawing to the service database; the feature selection module performs feature selection from the external database according to the package type characteristics of the electronic part pattern; the data integration module performs data preprocessing and normalization on the selected feature values to To obtain the data to be classified; and the service database to obtain the classification results of the packaging types of the electronic parts. According to the electronic component packaging classification system described in item 2 of the scope of the patent application, it further includes a training module and a parameter storage module. The training module is connected to the data integration module and the service database, and determines the training. The training scale and neural network parameters of the data set for training are used as the basis for subsequent classification. Among them, the convergence condition of training is that the cumulative error after the current training is less than the given threshold value, that is, the training is stopped, and the parameter storage module Connected to the training module and the service database to record training parameter data used by the training module. According to the electronic component packaging classification system described in item 3 of the scope of patent application, wherein the data integration module normalizes the feature values to the v _a and v _b intervals, satisfying , V _a <v _b relation, where, v 'is normalized to the feature value of v _a, v _b, v is the need for the normalized feature values, v _max is the maximum eigenvalue of a feature, and v _min is the minimum eigenvalue of a feature. According to the electronic component packaging classification system system described in item 4 of the scope of patent application, wherein the neural network parameters are convergence conditions, number of hidden layer neurons, number of hidden layers, initial learning rate, initial momentum, threshold value, Any one or a combination of the above, such as weight value and partial weight value. According to the electronic component packaging classification system described in item 5 of the scope of patent application, wherein the number of neurons in the hidden layer j satisfies ( x × ( input + output )), 1.5 < x <2, where input is the input package type There are 19 features, and output is the package type of the classification output. According to the electronic component package classification system described in item 6 of the scope of the patent application, the package type data records electronic component appearance information, printed circuit board restricted area information, solder joint information, geometric shape parameters, applicable field parameters, electrical Parameters, contact parameters, etc. any of the above or a combination thereof. According to the electronic component package classification system described in item 7 of the scope of patent application, the package type characteristics include the physical appearance of the electronic component, the physical pins of the electronic component, and the electronic component drawing. According to the electronic component packaging classification system described in item 8 of the scope of the patent application, wherein the weight ratio of the package type feature is that the electronic component pattern is larger than the physical appearance of the electronic component, and the physical appearance of the electronic component is greater than the physical pin of the electronic component. The electronic component package classification system according to item 9 of the scope of the patent application, wherein the physical appearance of the electronic component, the physical pins of the electronic component, and the electronic component pattern are selected from the number of pins of the electronic component, and the physical length of the original electronic component. , The largest electronic part entity length, the smallest electronic part entity length, the original electronic part entity width, the largest electronic part entity width, the smallest electronic part entity width, the electronic part entity height, the distance between the electronic part entity and the circuit board, Large electronic part pin length, small electronic part pin length, large electronic part pin width, small electronic part pin width, large electronic part pattern pin length, small electronic part pattern pin length, There are 19 types, such as large electronic part pattern pin width, small electronic part pattern pin width, electronic part pattern pin pitch X axis direction, and electronic part pattern pin pitch Y axis direction.