TW202025005A - A model-based machine learning system - Google Patents

Abstract

A model-based machine learning system for calculating optimum molding conditions includes a data storage device providing a set of training data; an injection molding process emulator producing a set of emulated sensing data according to molding conditions as inputted; an injection molding process state observation unit, determining an injection molding process state from molding conditions, sensing data and a quality state, wherein the quality state at least includes an acceptance state; and an injection molding process optimization unit including an injection molding condition optimizer, wherein a molding condition optimization model constructed in the injection molding condition optimizer is trained according to the injection molding process state as determined, and the molding condition optimization model after training is introduced into an injection molding production line.

Description

Model-based machine learning system

The present invention relates to a learning system, and particularly relates to a model-based machine learning system to calculate the best molding conditions for injection molding.

Injection molding is a complicated process. Taking plastic injection molding as an example, it is the result of a series of steps including plasticization of polymer materials, pressure injection into the membrane cavity, compression, cooling, and ejection. There are many factors that affect the quality of injection molding. In practice, from the initial mold trial to stable mass production, plastic injection products need to undergo a series of molding parameter tests and adjustments to confirm that the parameters can stably produce injection parts that meet the product design specifications. Even if the molding parameters have been adjusted, the molding quality will vary due to variations in the production environment. At this stage, in practice, the "experience" of personnel is mostly used to adjust and optimize the molding parameters to stabilize the molding quality. However, the parameters adjustment methods are different, the training of experienced adjusters is not easy, the adjusters' learning curve for new injection molding equipment and other factors, coupled with the shortcomings of high personnel costs and difficult quality control, how to overcome these difficulties? It is actually one of the important projects that need to be solved urgently in the molding manufacturing industry.

In practice, the main problems faced by the molding manufacturing industry include increasingly complex product designs, shrinking molding windows, increased quality of finished products affected by the molding environment, and decreased molding stability and yield. Moreover, the degree of product customization is increasing nowadays, and a small number of diverse manufacturing methods have resulted in an increase in the frequency of production lines. A lot of manpower is required to optimize the molding parameters to stabilize the quality of the finished product, so the labor cost is greatly increased.

Taking the traditional injection molding process as an example, the problems encountered in the optimization method of molding parameters, such as the difficulty of optimizing multiple acceptance conditions at the same time (the more complex the product design, the smaller the molding window, and the more acceptance conditions). It is preset to easily obtain the marking data of forming and quantitative quality, but the marking data is not easy to collect. The difficulties encountered in the traditional injection molding process include: it is difficult to evaluate the pros and cons of molding parameters, and even experienced molding process engineers cannot confirm the pros and cons of molding parameters by experience. Furthermore, a small number of diverse manufacturing production trends also make it difficult to effectively accumulate a large number of samples to support traditional machine learning methods. In addition, it is not easy to directly measure the quality indicators of injection products, such as the degree of burrs and the degree of warpage. Even if the molding parameters have been adjusted, the molding quality will vary due to variations in the production environment. At the present stage, we rely on experienced personnel to re-adjust and optimize the molding parameters. There are also problems such as high personnel costs and difficult quality control.

The present invention relates to a model-based machine learning system, which adopts artificial intelligence technology to construct a production environment variation model from historical data, and automatically optimizes molding process parameters, which can instantly compensate for quality variation caused by environmental variation.

According to one embodiment, a model-based machine learning system is proposed to calculate the best molding conditions for injection molding, including a data storage device for storing and processing data and providing a training data set; an injection molding program simulation The unit generates a set of simulated sensing data according to the input molding parameters; an injection molding program state monitoring unit, based on the molding parameters, the set of simulated sensing data, and a molding quality state to form an injection molding process environment state, where The molding quality status includes at least a good product identification result; and an injection molding process optimization unit, which uses a molding parameter optimizer to train a molding parameter optimization model based on the environment of the injection molding process, and optimize the molding parameters after training The model is introduced into an injection molding production line.

In order to have a better understanding of the above-mentioned and other aspects of the present invention, the following specific examples are given in conjunction with the accompanying drawings to describe in detail as follows:

In the embodiment of the present disclosure, a model-based machine learning system is proposed to calculate the optimal molding conditions for injection molding, so as to solve the difficulty in evaluating the merits of molding parameters encountered in the process of optimizing molding parameters in practice, and Processing the big data required in the training phase of artificial intelligence technology can also consider the quality of the molded product in real time, and quickly make real-time optimization. Furthermore, the model-based machine learning system proposed in the embodiment (including the simulation unit, monitoring unit, optimization unit described in the text, and the estimator, generator, inference, discriminator, and selector included in the aforementioned units, respectively) , Optimizer, etc.), which can be performed through one or more logical operation units and/or processors. Examples of the applicable logic operation unit and/or processor include (but not limited to) one or a combination of one or more of a chip, a circuit, a circuit board, and a recording medium storing an array of program codes. Please refer to Figure 1, which shows a block diagram of a model-based machine learning system. According to an embodiment of the present invention, an environment 10 (for example, actual injection molding data) is provided through a pre-designed experiment, so that an environment emulator 14 obtains an appropriate amount of marks from the environment 10 Data, and construct a running environment model; by interacting with the environment model, the optimization agent 11 of the system can still complete the optimization learning process under the situation that the environment interaction cannot be obtained in real time. Furthermore, the optimizer that has initially completed the learning in the optimization program 11 can still be updated in a timely manner based on the actual data accumulated in the actual injection molding process, so as to interact with the environment 10 in real time for reinforcement learning. Therefore, with the model-based machine learning system of the present invention, in addition to adjusting the molding process parameters of injection molding, it can also compensate for the quality or feature value variation caused by the variation of the molding environment, thereby optimizing and stabilizing the molding quality of the injection parts, and The learning of the optimizer can be strengthened according to the actual data accumulated by the injection parts according to the application requirements to re-optimize the molding parameters.

The following are related embodiments, together with the figures, to illustrate the system proposed in this disclosure in detail. However, the present disclosure is not limited to the units or devices exemplified in the system including the embodiments. Therefore, this disclosure does not show all possible embodiments, and other implementation aspects not mentioned in this disclosure may also be applicable. Those in the relevant field can change and modify the system of the embodiment without departing from the spirit and scope of the present disclosure to meet the needs of practical applications. Therefore, the contents of the description and illustrations are only used to describe the embodiments, not to limit the protection scope of the disclosure.

Figure 2 is a block diagram of model building and learning in a model-based machine learning system according to an embodiment of the present invention. Furthermore, the system block shown in Figure 2 can correspond to the flow of interaction between the environment simulator 14 and the optimization program 11 in Figure 1.

As shown in Figure 2, a model-based machine learning system for calculating the optimal molding conditions of injection molding includes a data storage device 10 _DS , an injection molding process emulator 400 (Injection molding process emulator), and an injection molding process emulator. An injection molding process state observation unit 200 (Injection molding process state observation unit) and an injection molding process optimization unit 300 are provided.

In the embodiment, a data storage device (Data storage) 10 _DS is used to store and process data, wherein the stored production raw data (Production Raw Data) undergoes data preprocessing 101 to provide a training data set D _TD . In one example, the production raw data includes, for example, the actual production mold number of injection molding, actual molding conditions, actual sensing data, and actual product quality status. The product quality status includes, for example, the classification results of good products (for example, True/False to indicate the classification of product quality), and quality data of each acceptance condition. In one example, the quality data of the acceptance condition is, for example, quantitative data such as injection molding flash, warpage, weight, and size. Furthermore, in an example, the data pre-processing 101 may include (but is not limited to) tools such as data filtering, data merging, and normalization.

In an embodiment, the injection molding process simulation unit 400 generates a set of emulated sensing data D _ES according to the input molding parameters of the molding condition.

In an embodiment, the injection molding process state monitoring unit 200 forms an injection molding process state (injection molding process state) S according to the molding parameters MC, analog sensing data D _ES, and molding quality state (QS). _k , where the molding quality status includes at least a good product identification result.

The injection molding program optimization unit 300 uses a molding parameter optimizer 310 (Injection molding condition optimizer) based on a reinforcement learning algorithm. The molding parameter optimizer 310 optimizes a molding parameter of the structure according to the injection molding process environment state S _k Model (Molding condition optimization model) for training. The trained molding parameter optimization model can be learned offline or imported into an injection molding production line for online learning.

The following is a further example of the injection molding program simulation unit 400, the injection molding program state monitoring unit 200, and the injection molding program optimization unit 300.

In one example, the injection molding process state monitoring unit 200 includes at least a qualitative molding quality inference engine 240 (Acceptance state inference engine), and the qualitative molding quality inference engine 240 creates a good product state classification model based on the training data set D _TD (Acceptance state classification model). The qualitative molding quality inference device 240 infers the simulated sensing data D _ES generated by the injection molding process simulation unit 400 according to the good product status classification model to infer the qualitative quality of the molded product with the set of simulated sensing data D _ES . At this time, the molding quality status collected by the injection molding process status monitoring unit 200 includes the good product identification result from the inference result of the qualitative molding quality inferring device 240. In the example, the qualitative molding quality inference device 240 can be (but not limitedly) updated after each production mold.

Of course, in other examples, the injection molding process status monitoring unit 200 may also include a quantitative or other qualitative inferring device or selector, and the model built in the inferring device or selector can, for example, infer a quantitative quality result ( For example, the quantitative molding quality inferring device 230 proposed in the following example), or a qualitative result that can be judged according to the quantitative quality result of the inference (such as the quality acceptance discriminator 250 proposed in the following example, or the qualitative result Molding quality data source selector 270). Common built-in inferential models are, for example, using Support Vector Classifier, Linear Discriminant, Nearest Neighbors, Decision Tree, Random Forest, Methods such as neural network (Neural Network) perform classification, but are not limited to the above categories.

In one example, the injection molding process state monitoring unit 200 further includes a qualitative molding quality data source selector 270 (Acceptance state input selector). The qualitative molding quality data source selector 270 has a built-in good product identification inference model, which determines that the qualitative molding quality inference device 240 deduces the set of simulated sensing data D _ES and its corresponding molded product belongs to the qualitative quality Good or bad products, and produce qualitative quality results of molding projects.

Furthermore, in an example, the injection molding process status monitoring unit 200 further includes a quantitative molding quality inference engine 230 (Molding quality inference engine), so that the molding quality status gathered by the injection molding process status monitoring unit (200) is in addition to good products The identification result (at least from the result deduced by the qualitative molding quality inferr 240), and further includes a quantitative quality result (Molding quality) deduced from the quantitative molding quality inferr 230. In the example, the quantitative molding quality inference device 230 can be (but not limitedly) updated after each production mold.

In one example, the quantitative molding quality inference unit 230 establishes a molding quality inference model based on the training data set. The quantitative molding quality inference device 230 can infer the quantitative quality results of the molded product with the set of simulated sensing data D _ES based on the quantitative quality inference model of the molding project and the simulated sensing data D _ES generated by the injection molding program simulation unit 400.

In one example, the injection molding process state monitoring unit 200 may further include a quality acceptance state identifier 250 (Acceptance state identifier) and a quantitative molding quality inferring device 230. The quantified quality results are used for quality judgment. For example, if the burr value (quantified quality result) of the molded product deduced by the quantitative molding quality inference device 230 is greater than 2mm, the quality acceptance discriminator 250 determines that the quantitative value of this item does not meet the acceptance conditions, and is less than or equal to 2mm. It is judged to meet the acceptance conditions. The quality acceptance discriminator 250 can simultaneously set acceptance conditions for many different quantitative items. Therefore, the quality acceptance discriminator 250 makes a qualitative determination based on the quantification result (using the quantification result to infer the qualitative result). Furthermore, the quality acceptance discriminator 250 can transmit the determined qualitative result to the qualitative molding quality data source selector 270, and the qualitative molding quality data source selector 270 corresponds to the quality determination result of the molded product being qualitative The quality is good or bad. Therefore, in this example, the molding quality status collected by the injection molding process status monitoring unit 200, the qualitative part of the good product identification result can be derived from the inferred result of the qualitative molding quality inference device 240 and the quality acceptance discriminator 250 makes a qualitative judgment based on the quantitative result (inference of the quantitative molding quality inferr 230), and selects it by the qualitative molding quality data source selector 270.

Furthermore, in an example, the injection molding process state monitoring unit 200 may further include a module 280 coupled to the qualitative molding quality data source selector 270 and the molding parameter optimizer 310, respectively. After the inference, the quantitative quality results of the molding items of the molded product deduced by the quantitative molding quality inferr 230, such as the molding item measurement data MQ (Molding quality), are entered into the module 280 for integration. The qualitative quality results of the molding items generated by the qualitative molding quality data source selector 270 after the qualitative results inferred from the qualitative molding quality inference device 240, such as the acceptance state (AS), are also Enter the module 280 for compilation.

In an embodiment, the injection molding program simulation unit 400 can construct a correlation model between molding parameters and sensing data based on the historical data of the actual manufacturing process, and simulate various sensing data of the mold based on the molding parameters input from the mold. Output.

In one example, the injection molding program simulation unit 400 can simulate the simulated sensing data (emulated sensing data) that is not actual data based on the relevant parameters and data distribution of the training data set (actual data), and according to the input molding parameter MC (molding condition). sensing data) D _ES . Therefore, the configuration of the injection molding program simulation unit 400 of the embodiment can make the qualitative molding quality inferr 240, or the combination of the qualitative molding quality inferr 240 and the quantitative molding quality inferr 230, not only for the training data set (actual data) A qualitative inference, or a qualitative and quantitative inference is performed, and a qualitative, qualitative and quantitative inference is also performed on the simulated sensing data D _ES (not actual data) generated by the injection molding program simulation unit 400. Therefore, the configuration of the injection molding program simulation unit 400 of the embodiment can increase the data content (including actual or non-real data) obtained by the injection molding program status monitoring unit 200. The following is an example of one implementation of the injection molding program simulation unit 400, but the disclosure is not limited thereto.

In one example, the injection molding program simulation unit 400 includes a statistical parameter estimator 410 and a random number generator 420. The parameter statistical estimator 410 can construct a correlation model according to the actual shaping parameters in the training data set and the individual actual sensing data distribution. For example, according to the actual molding parameters in the training data set and the statistics of the individual actual sensing data distribution, according to the simulated molding parameters MC input to the molding conditions of the injection molding program simulation unit 400, the individual corresponding to the simulated molding parameters is estimated. Statistics that simulate the distribution of sensing data. The estimation method is, for example, an interpolation method (using, for example, nearest neighbor interpolation, linear interpolation, cubic interpolation or Cubic or Cubic Spline), or estimation by other methods. In an example, the aforementioned statistics based on the actual individual sensing data distribution can be the average value m and standard deviation s of each data statistics; then the simulation data can be estimated by appropriate estimation methods such as interpolation or other methods , And simulate to estimate the statistics of individual simulated sensing data distribution, including the mean m and standard deviation s of each data statistics.

The random data simulation generator 420 is based on the aforementioned correlation model and can randomly generate a plurality of corresponding individual simulation sensing data according to the simulation molding parameters of the molding conditions input to the injection molding program simulation unit 400. Combining the corresponding individual analog sensing data can form a set of analog sensing data D _ES to provide to the injection molding process state monitoring unit (200). The random data simulation generator 420 randomly generates a plurality of corresponding individual simulated sensing data (such as simulated filling time) based on the estimated statistics of the individual sensing data distribution. Among them, the same simulated molding parameter generates multiple different simulated sensing data for the same sensing data item.

Therefore, the input and output simulated by the injection molding program simulation unit 400 have a one-to-many correspondence, that is, the same molding parameter corresponds to the same sensing data item, and different simulated sensing values will be generated. In the injection molding program simulation unit 400 proposed in the embodiment, the one-to-many correspondence between analog input and output achieved by it is in line with the actual injection molding process. In the actual injection molding process, although the process parameters are the same, different sensing data (such as device sensing data and internal sensing features of the mold) may also be generated.

According to the above example, the input molding parameter MC, the simulated sensing data D _ES , the molding project measurement data MQ (quantitative quality result of the molding project) inferred from the quantitative molding quality inferr 230, and the good product identification result AS (molding project The qualitative quality result can be derived from the qualitative molding quality inferring device 240 and the quality acceptance discriminator 250 and passing the qualitative molding quality data source selector 270), which can enter the module of the injection molding process state monitoring unit 200 280 for compilation. The data deduced by the qualitative molding quality inferring device 240 and the quality acceptance discriminator 250 may include the simulated sensing data D _ES (non-subject) from the training data set (historical data of the actual process) and the injection molding program simulation unit 400, respectively. Historical data of the actual manufacturing process).

Furthermore, in an example, the module 280 can also be used as a trigger of the molding parameter optimizer. If the qualitative molding quality data source selector 270 determines that the molded product of the set of simulated sensing data is a good product in terms of qualitative quality, the module 280 outputs the state of the injection molding process according to the mold order (such as outputting the mold second The injection molding process environment state S _k to the molding parameter optimizer 310) to complete the last training in the training process of the molding parameter optimization model of this round, and randomly select a set of initial molding parameters for the next round of molding parameter optimization model During the training process, the molding parameter optimizer 310 continues to train the molding parameter optimization model. After the production line operates for a period of time, the actual product quality results can be produced, or a predetermined time can be set to update the injection molding program simulation unit 400 and the injection molding program status monitoring unit 200.

If the molding parameter optimization model has not completed the training process of the molding parameter optimization model for this round (that is, the module 280 is driven to continue the optimization process), the molding parameter optimizer 310 can continue to perform according to the state of the injection molding process of the mold. The training process of the optimization model for the shaping parameters of this round. In one example, the molding parameter optimizer 310 can update the molding parameter optimization model according to the molding parameter optimization model of the mold and the simulated injection molding process environment state S _k gathered by the injection molding program status monitoring unit 200, and then Recommend and input another composition condition to the injection molding program simulation unit 400 to perform program simulation (through the injection molding program simulation unit 400) and program status monitoring (using the injection molding program status monitoring unit 200) , Until a set of optimized molding parameters are generated to complete the training process of the molding parameter optimization model for this round. The details have been detailed as above and will not be repeated here.

In the text, the training process of completing each round of the molding parameter optimization model means that the molding parameter optimizer 310 initially assigns a set of molding parameters to the injection molding program simulation unit 400, and then performs a parameter optimization process based on the existing parameter optimization model. , If the module 280 determines that this composition parameter will produce defective products, it recommends and inputs the molding parameters of another composition condition to the injection molding program simulation unit 400 until the recommended molding parameters enable the module 280 to determine that a simulated good product can be generated So far, the parameter optimization model training of this round is completed. Subsequently, the molding parameter optimizer 310 selects another set of new molding parameters to perform the next round of parameter optimization model training. At the beginning, the training process of completing a round of parameter optimization model may require many times, for example, 20 times of recommending and adjusting the molding parameters, before the module 280 can determine that a simulated good product can be produced. However, as the number of training rounds increases, the number of parameter adjustments required to complete each round will gradually decrease (that is, the number of parameter adjustments for each round to complete training gradually converges), because the training rounds have been learned from the past training rounds for shooting How to adjust the corresponding molding parameters in the molding state.

In addition, the user can selectively set the molding parameter optimization model training of the molding parameter optimizer 310 according to the actual needs of the application, and the training of the molding parameter optimization model has been initially completed, and can be imported into the actual injection molding production line for use. For example, it can be set in the total number of consecutive rounds R. The number of parameter adjustments completed in each round of training converges to a maximum of m times and reaches n% or more of the total number of rounds, which can be regarded as the preliminary completion of the shaping parameter optimization model. training. R is for example 10, 15, 20, 25, 30 rounds, or other rounds deemed suitable by the user; m is for example 5, 4, 3, or other suitable positive integer values, n% is for example 80%, 85 %, 90%, 95%, or other suitable ratio values, the present disclosure does not particularly limit the values of R, m, and n. Taking R=20, m=5, n%=95% as an example, it means that if 20 rounds of continuous training of the shaping parameter optimization model are performed, the number of parameter adjustments after each round of training converges to a maximum of 5 rounds. 95% or more of the total number of rounds, that is, there are 19 rounds of parameter adjustments up to 5 times (for example, rounds including 5, 4, 3, 2, and 1 rounds are included in the calculation). In order to preliminarily complete the training of the molding parameter optimization model of the embodiment, it can be imported into the actual injection molding production line for use.

According to an embodiment, the molding parameter optimizer 310 of the injection molding process optimization unit 300 constructs a molding parameter optimization model that includes at least one molding process state and multiple sets of corresponding molding parameter adjustment behaviors, wherein the sets correspond to The relationship is that the expected output of the corresponding molding parameter adjustment behaviors achieves the expected value of acceptable molding quality under at least the input molding process state. The molding parameter optimization model of the embodiment that can recommend and optimize the molding parameters is, for example, a neural network. Furthermore, the constructed optimization model of molding parameters can be updated automatically or by a user as needed, and the update frequency is not limited, and can be updated regularly or irregularly. This disclosure does not limit this .

In summary, in the embodiment, the training of the molding parameter optimizer 310 can be performed based on the good product status classification model and the molding item quality predictor model. Furthermore, in the embodiment, a reward evaluation R (Reward Evaluation) (for example, a reward evaluation unit RE) may be proposed to the injection molding program optimization unit 300 according to the molding quality status gathered by the injection molding program status monitoring unit 200. When the molding quality status is a good product (that is, the good product quality status is true), the reward is, for example, +1; when the molding quality status is a defective product (that is, the good product quality status is no), the reward is 0 or -1.

Figure 3 is a block diagram of a model-based machine learning system for online learning according to an embodiment of the present invention. Furthermore, the system block shown in Figure 3 can correspond to the flow of the optimization program 11 in Figure 1 on the environment 10.

As shown in FIG. 3, the injection molding production line introduced by the model-based machine learning system of the embodiment includes an actual injection molding manufacturing process (Actual injection molding process) 100. The recommended molding parameters MC _R (recommended molding conditions, a combination of multiple molding parameters) from the molding parameter optimizer 310 can be input into the actual injection molding manufacturing program 100, and the actual injection molding manufacturing program 100 is the actual molding parameter of the output device (applied molding condition) MC _A and actual sensing data D _{SD are sent} to the injection molding program state monitoring unit 200 and stored in a data storage device (Data storage) 10 _DS . The molding conditions (combination of multiple molding parameters) and the sensing data have a sequential sequence and a causal relationship, that is, the molding parameters are the cause (the timing is first), and the sensing data is the result (the timing is later). The sensing data can be divided into device sensing related data, including molding device sensing data, peripheral device sensing data, mold sensing data, and so on.

In one embodiment, the actual injection molding manufacturing process 100 is completed by a molding environment 110 including a series of actions such as molding parameter settings, injection actions, and output of molded products. The molding environment 110 includes molding equipment, molds, and related peripheral equipment. Or auxiliary systems, such as mold temperature controllers, dryers, cooling systems, etc.

Furthermore, as described above with regard to the state monitoring unit 200 for the injection molding process, it may at least include a qualitative molding quality inference device (240). As shown in Figure 3, the qualitative molding quality inference device 240 is based on the established good product status classification model (based on a training data set or an updated data set) for the actual sensing data output by the actual injection molding manufacturing process 100 D _SD makes an inference to infer the qualitative quality of the molded product with the actual sensing data D _SD . In turn, the injection molding process status monitoring unit 200 can summarize the molding quality status including at least one identification result of a good product for the actual injection molding manufacturing process. Or, as mentioned above, the injection molding process status monitoring unit 200 may also include a quantitative inferring device (such as the quantitative molding quality inferring device 230 proposed in the above example), or other qualitative inferring devices (such as the basis proposed in the above example The quantified result is used to infer the quality acceptance discriminator 250) of the qualitative result. The injection molding process state monitoring unit 200 may also include a selector related to qualitative/quantification; as shown in Figure 3, the example selector includes, for example, a qualitative molding quality data source selector 270 and a quantitative molding quality data The source selector 260 selects and judges the results according to a plurality of different qualitative or quantitative forming quality data sources.

Furthermore, the model-based machine learning system of the embodiment can optionally further include a molded product inspection system 210 (sampling inspection update), which samples the actual products of the injection molding production line and performs the actual quantity of the sampled products. Measurement (for example, actual measurement on the hardware machine equipment corresponding to the quality item). The actual quality measurement results obtained by the molding product inspection system (for example, actual measurement data MQ _{M of} molding items) can be transmitted to the quantitative molding quality data source selector 260 of the injection molding process status monitoring unit 200. Therefore, in this example, the quantitative molding quality data source selector 260 aggregates the actual quality measurement results from the molding product inspection system 210, and the quantitative molding quality inferor 230 (which can be updated every mold time) for the actual sensing data D _SD quantitative quality inference results. Therefore, in an example, the quantitative molding quality data source selector 260 can determine the quality of the molded product with the actual molding quality data based on a plurality of quantitative molding quality data sources (such as the two sources shown in Figure 3) Quantify the quality, and send the molding project measurement data MQ (Molding quality) to the module 280. In one example, the priority order of the sources of the quantitative forming quality data source selector 260 is, for example, the forming product inspection system 210 or the quantitative forming quality inferring device 230. But this disclosure is not limited to this.

As in the foregoing example, the injection molding process status monitoring unit 200 may further include a quality acceptance discriminator 250, which qualitatively qualifies the actual quality measurement results and the quantitative quality inference results gathered from the quantitative molding quality data source selector 260 Quality judgment. Therefore, the quality acceptance discriminator 250 can be used to determine the molded products (the actual measurement results from the molded product inspection system 210) corresponding to the actual quality measurement results gathered by the quantitative molding quality data source selector 260 and the corresponding The molded products with the quantitative quality inference results (from the quantitative molding quality inferr 230) are good or defective in terms of qualitative quality.

Furthermore, the model-based machine learning system of the embodiment can optionally further include an external input unit 220, which can directly input the acceptance qualitative results judged by random inspection of the actual products of the injection molding production line; for example, the inspector can directly Observe and identify the acceptance or non-acceptance of these randomly inspected molded products, and directly input the judgment result into a processor. Therefore, the external input unit 220 can also be referred to as an external acceptance state input unit. In an example, the acceptance quality result of the external input unit 220 is transmitted to the quality molding quality data source selector 270 of the injection molding process status monitoring unit 200. Therefore, in an example, the qualitative molding quality data source selector 270 can determine the molding with the actual molding quality data based on multiple qualitative molding quality data sources (such as the three sources shown in Figure 3) The product is good or bad in terms of qualitative quality. As shown in Figure 3, the qualitative molding quality data source of the qualitative molding quality data source selector 270 is, for example, the external input unit 220 (which can be updated every mold time), and the qualitative molding quality inference device 240 (which can be updated every mold time). ) The inferred qualitative result and the qualitative result determined by the quality acceptance discriminator 250 (infer the qualitative result based on the quantified result), these results are selected and determined by the qualitative forming quality data source selector 270. The qualitative quality results (for example, the good product identification result AS) of the molding items generated after the qualitative molding quality data source selector 270 are entered into the module 280 for consolidation.

According to the embodiment, the good product identification is one of the quality of qualitative molding; in one example, the priority order of the source of the good product identification result AS is, for example, the external input unit 220, the qualitative molding quality inferr 240, or the quality acceptance discriminator 250. But this disclosure is not limited to this.

Therefore, as shown in Figure 3, the module 280 of the state monitoring unit 200 of the injection molding process integrates the actual molding parameters MC _A , the actual sensing data D _SD , and the molding project measurement data MQ (from the quantitative molding quality data source). Selector 260), good product identification result AS (from qualitative molding quality data source selector 270). Furthermore, in the example, the module 280 can also be used as the trigger of the molding parameter optimizer 310, that is, when the good product identification result AS is true (good product), the online learning of the molding parameter optimizer 310 ( Reinforce learning and optimize again); On the contrary, when the good product identification result AS is no (defective product), the molding parameters are optimized according to the state of the injection molding program of the mold.

More specifically, in an example, if the qualitative molding quality data source selector 270 infers according to the acceptable qualitative results of the external input unit 220, according to the qualitative molding quality inferr 240 and the quality acceptable discriminator 250 The aforementioned molded product corresponding to the actual molding quality data is determined to be a good product in terms of qualitative quality, then the module 280 stops triggering the molding parameter optimizer 310, and inputs the aforementioned recommended molding conditions (molding parameters) for actual injection molding manufacturing The next mold time of the program 100, and the molding parameter optimizer 310 performs incremental learning in batches on the injection molding production line.

If the qualitative molding quality data source selector 270 infers the aforementioned actual molding quality data based on the acceptance qualitative results of the external input unit 220, the qualitative molding quality inferr 240 and the quality acceptance discriminator 250 If the corresponding molded product is determined to be defective in terms of qualitative quality, the module 280 triggers the molding parameter optimizer 310 to optimize the molding parameters. The molding parameter optimizer 310 can perform incremental learning of the molding parameter optimization model according to the molding parameter optimization model and the injection molding process environment state S _k collected by the injection molding program status monitoring unit 200. The molding parameter optimizer 310 recommends and inputs another composition molding condition to the actual injection molding manufacturing program 100. Alternatively, depending on the actual application situation, the molding parameter optimizer 310 may perform the training of the molding parameter optimization model again as shown in FIG. 2 (the training related content is as described above).

In addition, in an example, the good product identification result AS can be displayed in the external input unit 220 in real time, and the user only needs to mark the predicted error result, which can reduce the user's operating load. Furthermore, the quantified molding quality data can also be displayed in the external input unit 220 in real time, and at the same time, the acceptance conditions input by the user are combined to automatically determine the good product identification to reduce the user's operating load.

In summary, like the embodiment system shown in FIG. 3, the injection molding process status monitoring unit 200 mainly aggregates the data produced by the actual injection molding manufacturing process 100 to describe the molding process environment status of the current mold. Therefore, the molding process environment state S _k includes the complete data produced by the actual injection molding manufacturing process 100, such as the actual molding parameters MC _A , the actual sensing data D _SD, and the quality status of the molded product. The molded product quality status may include at least quality Chemical indicators, or both qualitative indicators and quantitative indicators. In the embodiment, the qualitative index includes at least the good product identification result AS (based on the qualitative molding quality data source), and may also include the qualitative results of other injection molded products, such as whether there are flow marks, whether there are binary classifications such as jet lines. result. In the embodiment, the quantitative index includes the molding project measurement data MQ (based on the quantitative molding quality data source), such as the burr length of the injection molded product, the weight of the finished product, the size of the finished product, the degree of warpage, or the quantification of other factors that affect the product. data. The molding parameter optimizer 310 is the corresponding relationship between a molding program state and molding parameter adjustment behavior, determines the parameter adjustment behavior according to the input molding program state, and generates a set of optimized molding parameters for the next mold time. parameter. Furthermore, in an example, the history data of each mold molding manufacturing is stored in the data storage device (production data storage part) 10 _DS and is selectively synchronized to a centralized management system (Centralized Management System).

In addition, the forming parameter optimizer 310, the qualitative forming quality inferring device 240, and the quantitative forming quality inferring device 230 proposed in the embodiment all have their corresponding inference models, and they can be updated every mold time. The update mechanism is described as follows:

When the good product identification result is true (good product), the molding parameter optimizer 310 performs incremental learning of the molding parameter optimization model in batches according to the parameter adjustment data of the batch;

The quantitative molding quality inference unit 230 performs incremental learning of the quantitative quality inference model according to the actual quantitative quality measurement results when the model provides the actual quantitative quality measurement results; and

When the qualitative molding quality inference device 240 provides the actual qualitative quality measurement result for the model, it performs incremental learning of the qualitative quality inference model according to the actual qualitative quality measurement result.

Furthermore, the good product identification inference model of the qualitative molding quality data source selector 270 can perform incremental learning of the good product identification inference model based on the actual good product identification results when the model provides the actual good product identification results.

According to the above embodiment, the injection molding program simulation unit 400 can reduce the dependence of the parameter optimization learning process on actual data, improve the use efficiency of actual production data, and thereby improve the efficiency of parameter optimization learning (simulation vs. actual injection). Furthermore, compared with the molding parameter adjustment mode of the traditional process, the parameter adjustment mode of the molding parameter optimizer 310 of the embodiment can simultaneously adjust multiple molding parameters for multiple optimization targets (quality inspection items, acceptance conditions) to achieve molding The goal of parameter optimization is therefore a systematic and efficient parameter adjustment mode.

The injection molding process state monitoring unit 200 of the embodiment includes a molding quality inferring device (such as a qualitative molding quality inferring device 240 and a quantitative molding quality inferring device 230), which is a requirement for constructing the state of the injection molding process, which can reduce the need for marking data. And help determine the timing of parameter optimization (for example, the module 280 acts as a driver).

Based on the above, the embodiment proposes a model-based machine learning system, which uses the injection molding program simulation unit 400 to construct a correlation model (molding parameter optimization model) between the injection molding process environment state _Sk and the molding parameter adjustment behavior , And only a small amount of actual data is needed to build a molding parameter optimization model, which can greatly reduce the amount of actual data required for molding parameter optimization. Furthermore, as the number of training rounds of the shaping parameter optimizer 310 increases, the number of parameter adjustments required in each round will gradually decrease and converge to a very small number of times. Therefore, the system of the embodiment can quickly obtain the best molding conditions for injection molding. According to the experiment, the simulation result of the molding parameter optimizer 310 that has completed the learning shows that there is a 99.6% probability that the parameter optimization can be completed within 3 molds (Figure 3). Preliminary verification shows that the model-based molding parameter optimizer proposed in the embodiment can reduce the number of mold trials required for the molding parameter optimization process. Therefore, the traditional injection process is manually adjusted by experienced and high-energy operators and most of them adjust a single parameter at a time. For the injection molding device of the system of the application embodiment, multiple acceptance conditions can be optimized at the same time, which greatly reduces the search The process time of appropriate molding parameters can efficiently and instantly obtain multiple optimized molding parameters that meet the application conditions and requirements (for example, the material properties of the manufactured products are different, and the climatic conditions of the manufacturing place are different). When the example is applied to a molding process with complex product design (the smaller the molding window, the more acceptance conditions), the evaluation and confirmation of optimized molding parameters and the efficiency are significantly improved. Therefore, the system of the embodiment has extremely high economic value and benefits in industrial applications. In summary, the model-based machine learning system proposed in the embodiment can solve the difficulties encountered in evaluating the pros and cons of molding parameters in the process of optimizing molding parameters in practice, processing the big data required in the artificial intelligence technology training stage, and real-time Consider the molding quality (you can know the quality of the product in real time, and quickly optimize it in real time).

For example, the illustrated content in the embodiment is used to describe one of the embodiments or application examples of the present disclosure, and the present disclosure is not limited to the scope and aspect of the illustrated content. Other embodiments, such as combinations of different components or known components may be applicable, and can be adjusted and modified according to actual application requirements. Therefore, the content of the illustration is for illustrative purposes only, not for limitation. Generally the knowledgeable person should know that, applying the relevant system of the embodiment of the disclosure, the molding parameters, measurement data, sensing data, acceptance items to be optimized, etc., the selection of the details can be seen to affect the relevance of the actual application process This disclosure has no restrictions on making appropriate choices and adjustments based on factors.

In summary, although the present invention has been disclosed in the above embodiments, it is not intended to limit the present invention. Those with ordinary knowledge in the technical field of the present invention can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention shall be subject to those defined by the attached patent scope.

10 _DS : Data storage device 101: Data pre-processing D _TD : Training data set 100: Actual injection molding manufacturing program 110: Molding environment 200: Injection molding program status monitoring unit 210: Molded product inspection system 220: External input unit 230: Quantification Molding quality inferring device 240: Qualitative molding quality inferring device 250: Quality acceptance discriminator 260: Quantitative molding quality data source selector 270: Qualitative molding quality data source selector 280: Module MC: Molding parameter D _ES : Simulation Sensing data S _k : environment state of the injection molding process RE: reward evaluation unit R: reward evaluation MC _R : recommended molding parameters MQ _M : actual measurement data MC _A : actual molding parameters D _SD : actual sensing data 300: injection molding Program optimization unit 310: molding parameter optimizer 400: injection molding program simulation unit 410: parameter statistical estimator 420: random data simulation generator m: average value s: standard deviation

Figure 1 shows a block diagram of a model-based machine learning system. Figure 2 is a block diagram of model building and learning in a model-based machine learning system according to an embodiment of the present invention. Figure 3 is a block diagram of a model-based machine learning system for online learning according to an embodiment of the present invention.

10 _DS : Data storage device

101: Data pre-processing

D _TD : training data set

200: Injection molding program status monitoring unit

230: Quantitative molding quality inference device

240: qualitative forming quality inference device

250: Quality Acceptance Discriminator

270: Qualitative molding quality data source selector

280: Module

MC: Molding parameters

D _ES : analog sensing data

S _k : Environmental status of injection molding process

RE: Reward Evaluation Unit

R: reward evaluation

300: Injection molding program optimization unit

310: Molding parameter optimizer

400: Injection molding program simulation unit

410: Parameter Statistics Estimator

420: Random data simulation generator

m: average

σ: standard deviation

Claims (21)

A model-based machine learning system for calculating the best molding conditions for injection molding. The model-based machine learning system includes: a data storage device for storing and processing data, wherein the data storage device stores and processes a A training data set is provided after the original data; an injection molding program simulation unit, which generates a set of simulated sensing data according to the input molding parameters; an injection molding program state monitoring unit, based on the molding parameters, the set of simulated sensing data and A molding quality status is to constitute an injection molding process environment status, wherein the molding quality status includes at least a good product identification result; and an injection molding program optimization unit that uses a molding parameter based on a reinforcement learning algorithm An optimizer, the molding parameter optimizer trains a molding parameter optimization model according to the environment state of the injection molding process, and the trained molding parameter optimization model is imported into an injection molding production line. For the model-based machine learning system described in item 1 of the scope of patent application, the injection molding program simulation unit includes: a parameter statistical estimator, based on the actual molding parameters in the training data set and individual actual sensing data distribution, Constructing a correlation model; and a random data simulation generator, based on the correlation model, randomly generating a plurality of corresponding individual simulation sensing data according to the simulation molding parameters of the molding conditions input to the injection molding program simulation unit . The model-based machine learning system described in item 2 of the scope of patent application, wherein the parameter statistical estimator of the injection molding program simulation unit is based on the actual molding parameters and the actual individual sensing data in the training data set The statistics of the distribution are based on the simulated molding parameters input to the molding conditions of the injection molding program simulation unit, and the statistics of the individual simulated sensing data distribution corresponding to the simulated molding parameters are estimated. The statistics of the measured data distribution and the statistics of the individual simulated sensing data distributions each include the mean (m) and standard deviation (s) of the data statistics. According to the model-based machine learning system described in item 3 of the scope of patent application, the random data simulation generator randomly generates a plurality of corresponding individual simulation senses based on the estimated statistics of the aforementioned individual sensing data distribution Measurement data, wherein the same simulated molding parameter generates multiple different simulated sensing data for the same sensing data item. According to the model-based machine learning system described in item 1 of the scope of patent application, the state monitoring unit of the injection molding process includes a qualitative molding quality inferring device, and the qualitative molding quality inferring device establishes a good product based on the training data set The state classification model, the qualitative molding quality inference device is based on the good product state classification model to infer the set of simulated sensing data generated by the injection molding program simulation unit to infer the molding product with the set of simulated sensing data Qualitative quality. For the model-based machine learning system described in item 5 of the scope of patent application, wherein the injection molding process state monitoring unit further includes a qualitative molding quality data source selector, and the qualitative molding quality data source selector determines the quality The chemical molding quality inferring device infers that the corresponding molded product is a good product or a defective product in terms of qualitative quality after the set of simulated sensing data is deduced. According to the model-based machine learning system described in item 5 of the scope of patent application, wherein the injection molding process state monitoring unit further includes a quantitative molding quality inference device, and the molding quality state aggregated by the injection molding process state monitoring unit further includes A quantified quality result. The model-based machine learning system described in item 7 of the scope of patent application, wherein the quantitative molding quality inferring device establishes a molding item quantitative quality inference model based on the training data set, and the quantitative molding quality inferring device quantifies the molding item according to the training data set The quality inference model is compared with the set of simulated sensing data generated by the injection molding program simulation unit, and the quantitative quality result of the molded product with the set of simulated sensing data is inferred. The model-based machine learning system described in item 8 of the scope of patent application, wherein the injection molding process state monitoring unit further includes a quality acceptance discriminator and the quantitative molding quality inference device, and the quality acceptance discriminator is The quantitative quality result of the quantitative molding quality inferring device performs quality judgment. According to the model-based machine learning system described in item 9 of the scope of patent application, the injection molding process state monitoring unit further includes a qualitative molding quality data source selector and the quality acceptance discriminator to determine the corresponding quality The molded product that accepts the quality judgment result of the discriminator is good or bad in terms of qualitative quality. According to the model-based machine learning system described in item 6 of the scope of patent application, the injection molding process state monitoring unit further includes a module, respectively coupled to the qualitative molding quality data source selector and the molding parameter optimizer, Wherein, if the qualitative molding quality data source selector determines that the molded product of the set of simulated sensing data is a good product in terms of qualitative quality when the molding parameter optimization model is trained for one round, the module outputs the The injection molding process environment state is sent to the molding parameter optimizer, and the molding parameter optimizer randomly selects a set of initial molding parameters for the next round of training of the molding parameter optimization model; if the qualitative molding quality data is in this round The source selector determines that the molded product of the set of simulated sensing data is defective in qualitative quality, and the module is driven to make the molding parameter optimizer continue to train the molding parameter optimization model of the round. The model-based machine learning system described in the first item of the scope of patent application, wherein the molding parameter optimization model constructed by the molding parameter optimizer includes multiple sets of correspondences between at least one molding process state and the corresponding molding parameter adjustment behavior , The groups of correspondences are respectively the expected output of the corresponding molding parameter adjustment behaviors to achieve the expected value of acceptable molding quality under the input at least the molding process state. The model-based machine learning system described in the first item of the scope of patent application, wherein the injection molding production line includes an actual injection molding manufacturing process, and inputting the recommended molding conditions from the molding parameter optimizer in the actual injection molding manufacturing process, The actual injection molding manufacturing program outputs the actual molding parameters and actual sensing data of the equipment to the injection molding program state monitoring unit. The model-based machine learning system described in item 13 of the scope of patent application, wherein the injection molding process state monitoring unit includes a qualitative molding quality inferring device, which is classified according to the established state of a good product The model is used to infer the actual sensing data output by the actual injection molding manufacturing process to infer the qualitative quality of the molded product with the actual sensing data. According to the model-based machine learning system described in item 14 of the scope of patent application, the injection molding process state monitoring unit further includes a qualitative molding quality data source selector, and the qualitative molding quality data source selector determines the quality The chemical molding quality inference device infers that the molded product with the actual sensing data is good or defective in terms of qualitative quality. For example, the model-based machine learning system described in item 15 of the scope of patent application may further include an external input unit to input the acceptance qualitative results obtained by sampling the actual products of the injection molding production line. The qualitative result is sent to the qualitative forming quality data source selector. For example, the model-based machine learning system described in item 14 of the scope of patent application may further include a molded product inspection system for sampling and actual quality measurement of the actual products of the injection molding production line. The injection molding program status monitoring unit further includes A quantitative molding quality inferring device and a quantitative molding quality data source selector. The actual injection molding manufacturing process further outputs actual sensing data to the quantitative molding quality inferring device. The quantitative molding quality data source selector includes the molded product. The actual quality measurement result of the inspection system and the quantitative quality inference result of the quantitative molding quality inferring device of the actual sensing data. The model-based machine learning system described in item 17 of the scope of patent application, wherein the injection molding process state monitoring unit further includes a quality acceptance discriminator which selects the source of the quantitative molding quality data The actual quality measurement results and the quantitative quality inference results gathered by the device are used to make a qualitative quality judgment. The model-based machine learning system described in item 18 of the scope of patent application, wherein the injection molding process state monitoring unit further includes a qualitative molding quality data source selector, and the qualitative molding quality data source selector determines the corresponding The molded product with the result of the quality acceptance discriminator is good or defective in terms of qualitative quality. According to the model-based machine learning system described in item 19 of the scope of patent application, the injection molding process state monitoring unit further includes a module, respectively coupled to the qualitative molding quality data source selector and the molding parameter optimizer, Wherein, if the qualitative molding quality data source selector infers the molding product quality corresponding to the actual molding quality data according to the acceptance qualitative results of the external input unit and the qualitative molding quality inference device And according to the quality judgment result of the quality acceptance discriminator, it is determined to be a good product in terms of qualitative quality, then the module stops triggering the molding parameter optimizer, and enters the aforementioned recommended molding conditions in the next step of the actual injection molding manufacturing process Mold order; if the qualitative molding quality data source selector infers the molding products corresponding to the actual molding quality data according to the acceptance qualitative results of the external input unit and the qualitative molding quality inferring device If it is determined as a defective product based on the qualitative quality and the quality discrimination result of the quality acceptance discriminator, the module triggers the molding parameter optimizer to optimize the molding parameters. For the model-based machine learning system described in the first item of the scope of patent application, the optimized model of molding parameters after training is imported into the injection molding production line for online learning.

Images