TWI710920B - System and method for estimating product yield by artificial intelligence technology - Google Patents

Abstract

A system for estimating product yield by artificial intelligence technology, which includes: an information setting module for setting plural information of the machine; a measurement module for measuring A product further obtains a product information; a smart analysis database connected to the information setting module and used to receive the complex information, analyze the complex information to obtain decision information; and a comparison model Group, the comparison module is used to compare the decision information with the product information to obtain an analysis information of the product, and return the analysis information after the comparison to the intelligent analysis database.

Description

System and method for estimating product yield by artificial intelligence technology

The present invention relates to a system and method for estimating product yield, in particular to a system and method for artificial intelligence technology to estimate product yield.

For the traditional parameters collected by the machine, use statistical methods and coordinate with the method of product measurement to find out the most important factor affecting the product, and use this factor to estimate the product; when considering the largest factor of the product, it is also considered as a stand-alone machine, and there is no series connection. The entire production line ignores the correlation between the entire production line. When the environment changes greatly, the relevant environmental data is not collected, so it is not easy to find important factors for product estimation.

With the development and evolution of deep learning, some people use Reinforcement Learning to try to make machines imitate a series of human behaviors. The main concept is based on the current "state" of the "environment." Performing "actions" will get rewards for performing actions. Then due to rewards and actions, they will change the environment and state, trying to maximize their long-term rewards. However, enhanced learning technology is limited to single-machine replication. The situation of multiple sensors and heterogeneous machines is not considered.

Previously, when considering the maximum factor of the product, it was considered as a stand-alone machine, and there was The entire production line is connected in series, ignoring the correlation between the entire production line. When the environment changes greatly, the relevant environmental data is not collected. Therefore, it is not easy to find important factors for product estimation. The previous method The maximum factor of the product under consideration does not update the model parameters in real time as the machine ages, resulting in errors in product quality.

In view of the shortcomings of the above-mentioned conventional technology and at the same time improving the above-mentioned neglect of the correlation between the entire production line, the present invention proposes a system that uses artificial intelligence technology to estimate the product yield, which can solve the above-mentioned shortcomings through a supervised learning method.

In order to achieve the above objectives, a system for estimating product yield based on artificial intelligence technology proposed in the present invention includes: an information setting module for setting complex information of the machine; and a measurement model Group, the measurement module is used to measure a product to obtain product information; a smart analysis database, the smart analysis database is connected to the information setting module, and is used to receive the complex information, and perform analysis based on the complex information Then obtain a decision information; and a comparison module for comparing the decision information with the product information to obtain an analysis information of the product, and compare the analysis information after the comparison Return to the intelligent analysis database.

In the system for estimating product yield with artificial intelligence technology of the present invention, the intelligent analysis database further includes: a database module, the database module is used to receive the complex information, and the complex information store; A supervised learning module, the supervised learning module is connected to the database module, and receives the complex number information stored in the database module, and uses the complex number information for model training to obtain complex estimates; and A decision-making unit module, the decision-making unit module is connected to the supervised learning module, and receives the supervised learning module to obtain the complex number estimated value, and performs cross-analysis according to the complex number estimated value to obtain the decision information .

In the artificial intelligence technology estimation system of the present invention, the plural information further includes: a feed parameter, a production line machine information, and an environmental information.

In the system for estimating product yield by artificial intelligence technology of the present invention, the database module further includes: a relevance database and a non-relational database. The relevance database is suitable for a clear data definition. Related databases are more suitable for data with no fixed format.

In the artificial intelligence technology estimation system of the present invention, the supervised learning module further includes: a production line selection model unit, a training model unit, and a retraining time setting unit.

In the system for estimating product yield using artificial intelligence technology of the present invention, the selected production line model unit further includes: a data preprocessing unit, a loss function unit, a learning rate unit, a processor unit, and a model selection unit And a confidence index calculation unit.

The artificial intelligence technology of the present invention estimates the product yield system, wherein the processor unit is a plurality of processors for simultaneous training and test.

The present invention provides a method for estimating product yield by artificial intelligence technology, which includes: using an information setting module for setting plural information of a machine; transmitting the plural information to a database module, and performing the plural information Store, and the complex information is used for model training through a supervised learning module to obtain a complex estimated value. The complex estimated value is cross-analyzed by a decision unit module based on the complex estimated value to obtain a decision Information; and using a measurement module to measure a product to obtain product information, and compare the decision information with the product information through a comparison module to obtain an analysis information of the product.

In the method for estimating product yield by artificial intelligence technology of the present invention, the database module further includes: a relevance database and a non-relational database, the relevance database is suitable for clear data definition, and the non-relational database The database is more suitable for data with no fixed format.

In the method for estimating product yield using artificial intelligence technology of the present invention, the supervised learning module further includes: a production line selection model unit, a training model unit, and a retraining time setting unit.

In the traditional technology, the correlation between the entire production line and the change of the environment are ignored, so that the traditional method cannot effectively find the important factors that affect the product quality for product estimation. Therefore, the present invention uses a supervised learning method For product estimation, unlike traditional methods that use enhanced learning to select important factors, the supervised method uses the entire process production line to collect The machine parameters are used to train the model; and the present invention monitors the actual process parameters and setting parameters of the machine such as current and voltage during the machine process, as well as the environmental change parameters such as temperature and humidity in the environment, and collects data before the model training The various useful information related to the product or production is normalized to avoid the useful information being covered by other features due to the small value. In order to solve the variation of the process parameters of the machine growing over time, the present invention regards the trained model as a pre-trained model. Train the model, and then when the number of data reaches the set value, through the setting of retraining time and the processor unit in the training model function, a dual-track parallel system is adopted. One processor is trained and the other is tested. After the model has been trained by the training processor, the test processor will be notified to replace the model to achieve immediate replacement of the model and avoid the problem of prediction errors caused by the aging of the production line.

The above summary and the following detailed description and drawings are for the purpose of further explaining the methods, means and effects of this creation to achieve the intended purpose. The other purposes and advantages of this creation will be explained in the following description and diagrams.

1‧‧‧Information setting module

2‧‧‧Measurement Module

3‧‧‧Smart Analysis Database

4‧‧‧Comparison Module

31‧‧‧Database Module

32‧‧‧Supervised Learning Module

33‧‧‧Decision Unit Module

321‧‧‧Select production line model unit

322‧‧‧Training model unit

323‧‧‧Retraining time setting unit

The first figure is a schematic diagram of the system for estimating product yield by artificial intelligence technology of the present invention;

The second figure is a schematic diagram of the intelligent analysis database of the present invention.

The following is a specific example to illustrate the implementation of this creation In this way, people familiar with this technique can easily understand the advantages and effects of this creation from the content disclosed in this manual.

Please refer to the first figure, the schematic diagram of a system for inventing artificial intelligence technology to estimate product yield. A system for artificial intelligence technology to estimate product yield. The system for artificial intelligence technology to estimate product yield includes an information setting module. 1. The measurement module 2, the intelligent analysis database 3 and the comparison module 4; before the product manufacturing process, the information setting module 1 can be used to set the complex information of the machine, for example: when performing resistance welding After grinding, cleaning and drying the carrier of the bridge wire resistance welding, the bridge wire resistance welding is carried out, and then sent to the next process, and the information setting module 1 can be used before the bridge wire resistance welding product process The plural information used to set the machine, where the plural information includes feed parameters, production line machine information, and environmental information, where the feed parameters include: bridge wire sample number and thickness to be welded, and carrier for bridge wire resistance welding Material information and other parameters, the resistance welding production line machine information includes: resistance value, current, voltage, power, welding position, motor speed, wire tension, welding head pressure and other parameters, and environmental information includes: temperature, humidity and other parameters; When the product is processed, the processed product can be measured through the measurement module 2 to obtain product information, where the product information includes information such as tensile force, vibration test, electrical test value, etc.

Then, the intelligent analysis database 3 of the system for estimating product yield by artificial intelligence technology is connected to the information setting module 1, and is used to receive the complex information, analyze according to the complex information, and obtain decision information; and The comparison module 4 of the system for estimating the product yield rate of industrial intelligence technology is used to compare the decision information with the product information to obtain the analysis information of the product, and finally return the compared analysis information to The intelligent analysis database 3 is used to achieve real-time replacement models based on the analysis information and avoid the problem of forecast errors in the production line due to machine aging.

In another preferred embodiment, the intelligent analysis database 3 includes a database module 31, a supervised learning module 32, and a decision unit module 33, wherein the database module 31 is used to receive the complex information , And store the complex number information; the supervised learning module 32 is connected to the database module 31, and receives the complex number information stored by the database module 31, and uses the complex number information for model training to obtain the complex number Estimated value; and the decision unit module 33 is connected to the supervised learning module 32, and receives the supervised learning module 32 to obtain the complex estimated value, and performs cross-analysis according to the complex estimated value to obtain the decision News.

Among them, the database module 31 is used to record the real-time changes of the entire production line and to receive and store the plural information. The database module 31 can be divided into a relevance database and a non-relevance database. The sexual database is suitable for clear data definition, while the non-relational database is more suitable for data with no fixed format, usually stored in JavaScript Object Notation (JSON) files, so the use of the two databases can be based on the needs of the use situation Select the corresponding database to use.

Then, please refer to the second picture, and the second picture is based on the A schematic diagram of a system for estimating product yield using artificial intelligence technology further illustrates the schematic diagram of smart analysis database 3. After the complex information is stored, the complex information stored in the database module 31 is sent to the supervised learning module 32 for analysis, wherein the supervised learning module 32 includes a production line selection model unit 321 and a training model The unit 322 is composed of a retraining time setting unit 323; the retraining time setting unit 323 can be set according to the model training period (for example: 1 day, 15 days, 30 days, etc.) that the user wants to set. The number of pens (for example: 100 pens, 1000 pens, 10,000 pens, etc.) or after the actual measurement and model prediction accuracy is lower than a certain threshold (for example: the accuracy rate is lower than 90%, 95%, 99%, etc.) Therefore, before the complex information is sent to the supervised learning module 32 for analysis, the number of days of the model training period, the number of training sessions, and the ratio after the prediction accuracy rate is lower than a certain threshold can be performed.

Then, the user can perform analysis through the training model unit 322, and the training model unit 322 includes a data preprocessing unit, a loss function unit, a learning rate unit, a processor unit, a model selection unit, and a confidence index calculation unit. The data pre-processing unit normalizes the complex number information of the database module 31. The normalization method can use methods such as min-max standardization or z-score standardization. When min-max standardization is to perform linear transformation on the original data, In this way, the data falls into the [0,1] interval. For example, when the welding head pressure machine can be set from 0g to 650g, and the feedback current value is 0~1A, the difference between the two is too large. Although there is a difference in the weight of the operator, it cannot be balanced. Therefore, it is necessary to normalize the welding head pressure from 0g to 650g. Compressed to the [0,1] interval in equal proportions, and then throw all the values into machine learning for learning, where min is the minimum value of the sample data, max is the maximum value of the sample data; in addition, the z-score standardization is the data After processing, it conforms to the standard normal distribution, that is, data with a mean value of 0 and a standard deviation of 1. These two methods effectively prevent useful information from being obscured by other features due to small values.

Then, after the complex information is normalized, the loss function unit is used to select the loss function to evaluate the model, such as mean square error (MSE) or cross-entropy (Cross-entropy), and find better values through different loss functions. Good prediction accuracy; and the learning rate unit can choose the learning rate. By setting the learning rate, the model can be trained quickly. The learning rate is larger and the training speed is faster, but it is easy to fall into repetitive shocks. If the learning rate is small, the network is easy Falling into the local optimum and the training speed is slow, often used learning rate such as Adam, AdaGrad or Momentum.

Then, the processor unit can select hardware devices such as GPU, CPU, or TPU according to the user. The processor unit has at least two processors for simultaneous training and testing. It adopts a dual-track parallel system and one processor for training. , A processor is tested. After the model is trained by the training processor, a model file such as .tfrecords or .pkl will be generated and passed through the World Wide Web Application Programming Interface (WebAPI) Restart the test processor program and inform the location of the model file, and through the central processing unit (CPU), the model can be replaced within 1 second or the graphics processor can be used (Graphics Processing Unit, GPU) hardware equipment can be replaced within 10 seconds to complete the model replacement, so as to achieve real-time replacement of models and avoid the production line due to machine aging caused by forecast errors, in addition, the processor unit is also monitored Program management. When the monitoring program finds that there is only one processor left in the processor unit, it will give priority to setting this processor as the test processor, and the user can make relevant adjustments, such as when there is only one processor left Priority is set for training or testing processors, multi-processor allocation and other functions.

Then, the model selection unit can select the supervised learning model to be applied for evaluation, such as CNN, RNN, XGBoost... and other models can be selected; the confidence index calculation unit can use multiple models selected by the user to evaluate the confidence index and provide it to User reference; the function of selecting the production line model is to respond to the products produced by the existing production line. The user can choose the latest or the model with the highest accuracy in the past predictions for application, and contains small functions such as deletion, archiving, and backup; therefore, the supervision type After the learning module 32 receives the complex number information stored in the database module 31, the complex number information passes through the data preprocessing unit, loss function unit, learning rate unit, processor unit, model selection unit, and confidence index of the supervised learning model The computing unit and other units perform model training to obtain complex estimated values, which include confidence index unit values and model estimated values; the decision-making unit module 33 receives the supervised learning module 32 to obtain the complex estimated values , And perform cross-analysis according to the complex estimated value to obtain the decision information; finally, the comparison module 4 is used to compare the decision information with the product information to obtain an analysis information of the product, and compare the ratio The analyzed information after the analysis is returned to the intelligent analysis database 3, so as to achieve real-time replacement of the model based on the analysis information and avoid the problem of prediction errors caused by the aging of the production line.

For those who need to clarify, the above are only the preferred embodiments of this case, and are not used to limit the creation. If the changes made according to the conception of this creation are within the scope of the creation spirit, for example, regarding the configuration or layout The changes in the form, and the equivalent effects of various changes, modifications and applications, should be included in the scope of rights in this case, and shall be stated.

1‧‧‧Information setting module

2‧‧‧Measurement Module

3‧‧‧Smart Analysis Database

4‧‧‧Comparison Module

31‧‧‧Database Module

32‧‧‧Supervised Learning Module

33‧‧‧Decision Unit Module

Claims (8)

A system for estimating product yield by artificial intelligence technology, which includes: an information setting module for setting plural information of the machine; a measurement module for measuring A product obtains a product information; a smart analysis database, which is connected to the information setting module, receives the plural information through a database module of the smart analysis database, and processes the plural information Stored, the complex information is trained by a supervised learning module of the intelligent analysis database, wherein the model training is analyzed by a training model unit of the supervised learning module to obtain the complex estimated value, wherein the A processor unit of the training model unit has a complex number of processors for simultaneous training and testing. A dual-track parallel system is adopted. One processor is used for training and the other is used for testing. The complex estimated value is analyzed through the intelligence. A decision unit module of the library performs analysis to obtain decision information; and a comparison module for comparing the decision information with the product information to obtain an analysis information of the product, and The analyzed information after the comparison is returned to the intelligent analysis database. For example, the first artificial intelligence technology in the scope of the patent application is a system for estimating product yield, where the plural information includes: a feed parameter, a production line machine information, and an environmental information. For example, the first artificial intelligence technology in the scope of the patent application is a system for estimating product yield. The database module also includes: a relevance database and a non-relevance database. The relevance database is suitable for clear data By definition, the non-linked database is more suitable for data with no fixed format. For example, the first artificial intelligence technology in the scope of the patent application is a system for estimating product yield, in which the supervised learning module also includes: a production line selection model unit and a retraining time setting unit. For example, the first artificial intelligence technology in the scope of the patent application is a system for estimating product yield, where the training model unit includes: a data preprocessing unit, a loss function unit, a learning rate unit, a processor unit, and a model selection Unit and a confidence index calculation unit. An artificial intelligence technology method for estimating product yield, comprising: using an information setting module for setting plural information of a machine; transmitting the plural information to a database module, and storing the plural information, and The complex number information is trained by a training model unit of a supervised learning module to obtain complex number estimation values. A processor unit of the training model unit has a complex number processor for simultaneous training and testing. A dual-track parallel system is adopted. , One processor is training, and the other is testing, and the complex estimated value is cross-analyzed by a decision-making unit module based on the complex estimated value to obtain decision information; and A measurement module is used to measure a product to obtain product information, and a comparison module is used to compare the decision information with the product information to obtain an analysis information of the product. For example, the sixth item in the scope of the patent application is a method for estimating product yield using artificial intelligence technology, where the database module also includes: a relevance database and a non-relational database, the relevance database is suitable for clear data By definition, the non-linked database is more suitable for data with no fixed format. For example, item 6 of the scope of patent application is a method for estimating product yield using artificial intelligence technology, wherein the supervised learning module further includes: a production line selection model unit and a retraining time setting unit.

Images