A control method and device for a plunger system

Abstract

The application discloses a kind of injection system control method and device, belong to die casting machine control technical field, its method includes: with time sequence process parameter as input, future time period equipment parameter variation trend is output to build first model, with future time period equipment parameter variation trend as input, control parameter is output to build second model;Injection process is divided into multiple injection stages according to working state, after training first model and second model according to the historical process parameters, historical equipment parameters and historical control parameters of each injection stage, third model is obtained by merging;Real-time process parameters obtained are input into third model, third model is adjusted and optimized according to real-time equipment parameters, and target control parameter is obtained, and the real-time control parameter of parameter adjustment injection system and / or switch injection stage.The present application carries out prediction and self-adaptive adjustment to injection system according to different injection stages, controls the switching of injection stage, to realize the accurate control to injection system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of die casting machine control, and particularly relates to a control method and device for a die casting system. BACKGROUND

[0002] Die casting is a widely used high-efficiency forming method in non-ferrous metal product manufacturing, and has been widely used in the automobile, communication, and electronic industries. The die casting process of a die casting machine usually includes three stages: slow die casting, fast die casting, and pressure boosting and pressure maintaining. Each stage has different requirements for process parameters such as die casting speed and die casting pressure. The control of the die casting system of a die casting machine directly affects the filling process of the metal liquid, and further determines the size precision, surface and internal quality of the final die casting product. However, in actual production, due to changes in material types, temperature fluctuations, mold conditions and other process parameters, how to ensure the dynamic response and control accuracy of the die casting system under complex working conditions has become a major challenge in the field of die casting machine control systems.

[0003] Traditional die casting control systems mainly use PID control to adjust the die casting speed and pressure by manually setting the proportional, integral and derivative parameters (Kp, Ki and Kd). However, the traditional PID controller has problems such as response lag and overshoot when dealing with changes in working conditions, lacks self-adaptability, and is highly dependent on manual operation. To solve the limitations of traditional PID control, some intelligent control methods have emerged in recent years, such as fuzzy control and genetic algorithm optimization. These methods can improve the response ability and adaptability of the system to some extent, but they still face the following problems when applied to die casting machine die casting control: first, the prediction ability is limited, most existing intelligent control methods rely on real-time data feedback of the current working condition, lack the ability to predict future process states, and cannot predict possible process changes in advance, which can easily lead to control lag; second, real-time and adaptability are insufficient, although fuzzy control methods can optimize some process parameters, they often have difficulty in adjusting all relevant process parameters in real time and accurately during the die casting process, especially in the fast-changing die casting stage; third, lack of dynamic adjustment mechanism, existing intelligent control methods usually lack a dynamic adjustment mechanism for PID controller parameters, especially when the die casting requirements differ significantly in different stages (such as slow, fast, and pressure boosting and pressure maintaining) of the die casting process, it is difficult to optimize each stage.

[0004] Therefore, there is an urgent need for a die casting system control method and device to solve the technical problem of being unable to accurately control die casting parameters in the prior art. SUMMARY

[0005] Therefore, it is necessary to provide a control method and device for a injection system, which can predict and adaptively adjust the control parameters of the injection system in different injection stages, control the switching of the injection stages, and realize accurate control of the injection system.

[0006] To solve the above technical problems, in one aspect, the present application provides a control method for an injection system, comprising:

[0007] A first model is constructed with time sequence process parameters as input and future time period equipment parameter change trend as output, and a second model is constructed with future time period equipment parameter change trend as input and control parameter as output;

[0008] The injection process is divided into multiple injection stages according to working states, and the first model and the second model are trained according to historical process parameters, historical equipment parameters and historical control parameters of each injection stage, and then a third model is obtained by merging;

[0009] The obtained real-time process parameters are input into the third model, and the third model is adjusted and optimized according to real-time equipment parameters to obtain target control parameters, and the real-time control parameters of the injection system are adjusted and / or the injection stage is switched according to the target control parameters.

[0010] In one possible implementation, the first model is constructed with time sequence process parameters as input and future time period equipment parameter change trend as output, comprising:

[0011] The first model is constructed based on an LSTM neural network with time sequence process parameters as input and future time period equipment parameter change trend as output, and the first model comprises a first input layer, multiple LSTM layers, a full connection layer and a first output layer;

[0012] The input layer is used to receive historical time sequence process parameters;

[0013] The LSTM layer is used to process the historical time sequence process parameters, determine the dependency relationship of the time sequence in the time sequence process parameters, and extract time sequence features according to the dependency relationship;

[0014] The full connection layer is used to connect the LSTM layer and the output layer, map the time sequence features to obtain mapping features;

[0015] The output layer is used to output the change trend of the future time period equipment parameters according to the mapping features.

[0016] In one possible implementation, the second model is constructed with equipment parameter change trend as input and control parameter as output, comprising:

[0017] The second model is constructed based on the RBF neural network, with the device parameter change trend as input and the control parameter as output, and the second initial prediction model includes a second input layer, a plurality of hidden layers and a second output layer;

[0018] The second input layer is configured to receive the device parameter change trend.

[0019] The hidden layer is configured to extract feature information of the device parameter change trend.

[0020] The second output layer is configured to output the control parameter according to the feature information.

[0021] In a possible implementation manner, after the first model and the second model are trained according to the historical process parameters, the device parameters and the control parameters of each injection stage, a third model is obtained by merging, including:

[0022] The historical time-series process parameters and the historical time-series device parameters of different injection stages are obtained.

[0023] The first model is iteratively trained according to the historical time-series process parameters and the historical time-series device parameters, to obtain a first target model.

[0024] The historical device parameters and the historical control parameters of different injection stages are obtained.

[0025] The second model is iteratively trained according to the historical device parameters and the historical control parameters, to obtain a second target model.

[0026] The first target model and the second target model constitute the third model.

[0027] In a possible implementation manner, the first target model is obtained by iteratively training the first model according to the historical time-series process parameters and the historical time-series device parameters, including:

[0028] Based on the LSTM layer, the historical time-series process parameters are forward propagated to calculate the hidden state and the memory cell state of each time step.

[0029] The prediction value is generated by the fully connected layer according to the hidden state and the memory cell state, and the device parameter change trend in the future time period is output by the first output layer.

[0030] The first model is optimized according to a preset loss function, the historical time-series predicted device parameters and the device parameter change trend in the future time period, to obtain a first optimization model.

[0031] Based on an optimization algorithm, the output weight and the bias of the first optimization model are adjusted according to the gradient of the time step, to obtain the first target model.

[0032] In a possible implementation, the optimization algorithm is an Adam optimization algorithm, and the calculation formula of the output weight is:

[0033] ,

[0034] wherein, is the weight at time t, and are the momentum and the second moment estimation of the gradient respectively, is the learning rate, is a small constant to prevent the denominator from being zero.

[0035] In a possible implementation, the second model is iteratively trained according to the historical equipment parameters and the historical control parameters, to obtain a second target model, including:

[0036] The historical equipment parameters are clustered based on a K-means clustering algorithm, to obtain a plurality of cluster center points;

[0037] The hidden layer is activated based on a Gaussian function, and a width parameter of the cluster center point is calculated;

[0038] The weight of a second output layer is trained according to the width parameter, and a control parameter prediction result is output according to the weight;

[0039] The second model is iteratively trained based on a preset loss function and the control parameter prediction result, to obtain the second target model.

[0040] In a possible implementation, the third model is adjusted and optimized according to real-time equipment parameters, to obtain a target control parameter, including:

[0041] Real-time equipment parameters in a pressure injection process of the die casting machine are acquired;

[0042] The output result of the first model in the third model is adaptively optimized and adjusted according to the real-time equipment parameters;

[0043] The output result of the first model after the optimization and adjustment is taken as the input of the second model, to obtain the final target control parameter.

[0044] In a possible implementation, the pressure injection stage includes a slow pressure injection stage, a fast pressure injection stage and a pressure boosting and pressure maintaining stage, and the pressure injection stage of the pressure injection system is controlled according to the real-time control parameter, including:

[0045] When the real-time control parameter is in a preset slow pressure injection interval, the pressure injection system is controlled to be in the slow pressure injection stage;

[0046] When the real-time control parameter is in a preset rapid injection interval, the injection system is controlled to be in a rapid injection stage.

[0047] When the real-time control parameter is in a preset pressure increasing and maintaining interval, the injection system is controlled to be in a pressure increasing and maintaining stage.

[0048] When the real-time control parameter reaches a preset pressure increasing and maintaining upper limit, the injection system ends work.

[0049] In a second aspect, the application further provides an injection system control device, comprising:

[0050] a model construction module, configured to construct a first model with time sequence process parameters as input and future time period equipment parameter change trend as output, and construct a second model with future time period equipment parameter change trend as input and control parameter as output;

[0051] a model training module, configured to divide the injection process into multiple injection stages according to working state, and after training the first model and the second model according to historical process parameters, equipment parameters and control parameters of each injection stage, obtain a third model by merging;

[0052] an injection system control module, configured to obtain a target control parameter by adjusting and optimizing the third model according to real-time process parameters, and adjust real-time control parameters and / or switch injection stages according to the target control parameter.

[0053] The beneficial effects of the present invention are as follows: first, based on the LSTM neural network, a first time series model is constructed with time series process parameters as input and a future time period equipment parameter change trend as output, which can predict the equipment parameter change trend in the future period based on the historical time series process parameters and the current process parameters; based on the RBF neural network, the equipment parameter change trend in the future period is input and the control parameter is output, a second model is constructed, and the PID control parameters are further predicted based on the equipment parameters predicted by the LSTM; then, the injection molding process is divided into multiple injection stages according to the working state of the injection molding system, and the first model and the second model are iteratively trained according to the historical process parameters, equipment parameters and control parameters of each injection molding stage to obtain a third model for predicting the control parameters, and the model is trained in stages according to the parameter characteristics of different injection molding stages, so that the model can accurately predict the parameters of each stage; finally, according to the real-time process parameters and the trained third model, according to the real-time equipment parameters monitored in real time, the output results of the first model in the third model are adaptively adjusted and optimized, and finally the target control parameters are obtained, the real-time control parameters of the PID are adjusted according to the target control parameters, and the injection molding system is switched to different injection stages according to the adjusted real-time control parameters. The present invention constructs an LSTM-RBF fusion model, divides the shooting process of the shooting system into different shooting stages, predicts the control parameters of the PID controller according to real-time process parameters, and switches the working stage of the shooting system according to the control parameters. It realizes the accurate prediction of the control parameters of the PID controller and the adaptive adjustment of the control parameters and shooting stages of the shooting system, thereby improving the accuracy of the shooting system control. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0055] Figure 1 A schematic flow chart of an embodiment of the injection system control method provided by the present invention;

[0056] Figure 2 The present invention provides Figure 1 A flow chart of an embodiment of step S102;

[0057] Figure 3 The present invention provides Figure 2 A flow chart of an embodiment of step S202;

[0058] Figure 4 The present invention provides Figure 2Flowchart of the first embodiment of step S204;

[0059] Figure 5 Flowchart of the injection phase switching process provided by the present application;

[0060] Figure 6 Structure diagram of an embodiment of the injection system control device provided by the present application. DETAILED DESCRIPTION

[0061] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by a person skilled in the art without creative work fall within the protection scope of the present application.

[0062] In the description of the embodiments of the present application, unless otherwise specified, the meaning of "multiple" is two or more than two. The association relationship of the associated objects is described by "and / or", which means that there can be three relationships, for example: A and / or B, which can represent the three cases of A alone, A and B together, and B alone.

[0063] The "first", "second" and the like described in the embodiments of the present application are only for the purpose of description, and cannot be understood as indicating or implying the relative importance of the indicated technical features or implicitly indicating the number of the indicated technical features. Therefore, the technical features limited by "first" and "second" can explicitly or implicitly include at least one of the features.

[0064] In this document, the reference to "embodiments" means that the specific features, structures or properties described in connection with the embodiments can be included in at least one embodiment of the present application. The appearance of this phrase in various places in the specification does not necessarily mean that it refers to the same embodiment, nor is it independent or alternative to other embodiments. A person skilled in the art explicitly and implicitly understands that the embodiments described herein can be combined with other embodiments.

[0065] Before describing the embodiments, the related terms are explained:

[0066] The LSTM neural network is a kind of time recurrent neural network, which aims to solve the problem of gradient disappearance or explosion faced by traditional RNN when processing long sequence data, so as to capture long-term dependencies more effectively.

[0067] The RFB network is a network structure for target detection, which aims to enhance the feature extraction ability of the network by simulating the receptive field of human vision.

[0068] The application provides a injection system control method and device, which are described below.

[0069] Figure 1 An embodiment flowchart of the injection system control method provided by the application is shown in the figure. Figure 1 The injection system control method comprises the following steps.

[0070] S101, a first model is constructed with time sequence process parameters as input and future time period equipment parameter change trend as output, and a second model is constructed with future time period equipment parameter change trend as input and control parameter as output.

[0071] It should be noted that a time sequence model and a target detection model are constructed through a model management software on an electronic terminal, in this embodiment, the first time sequence model is constructed by using an LSTM neural network, and the target detection model is constructed by using an RBF neural network, but the application is not limited to the LSTM neural network and the RBF neural network, wherein the electronic terminal can realize data processing and model management functions, and can be a PC terminal, a mobile terminal, a portable terminal or the like.

[0072] It needs to be further explained that the timing process parameters refer to a series of key parameters that need to be set and adjusted in the die casting process, which determine the quality and generation efficiency of the die casting, including pressure parameters (injection pressure, pressure holding time, etc.), temperature parameters (melting temperature, mold temperature, casting cooling temperature, etc.), time parameters (injection time, cooling time, etc.) and speed parameters (injection speed, mold opening and closing speed, etc.) in different injection stages; the pressure parameters affect the filling and solidification process of the metal liquid, the temperature parameters affect the forming and cooling of the casting, the time parameters determine the time required in each stage of the die casting process, and the speed parameters affect the filling speed of the metal liquid and the opening and closing efficiency of the mold; the equipment parameters are the working state parameters of the die casting machine and its auxiliary equipment, such as the displacement amount of the servo cartridge valve core of the injection energy accumulator, the displacement amount of the servo cartridge valve core of the booster energy accumulator, and the pressure holding time; the control parameters are the parameters that need to be set in the control algorithm of the PID controller, which determine the response speed and stability of the control algorithm to the die casting process, including the proportional coefficient, the integral coefficient and the differential coefficient, among which, the proportional coefficient determines the response speed of the control algorithm to the deviation, the larger the proportional coefficient, the faster the response speed, but too large proportional coefficient will lead to overshoot and oscillation, the integral coefficient determines the response of the control algorithm to the cumulative effect of the deviation, the larger the integral coefficient, the more sensitive the system to the cumulative effect of the deviation, which helps to eliminate static error, but too large integral coefficient will lead to integral saturation and response lag, the differential coefficient determines the response of the control algorithm to the rate of change of the deviation, the larger the differential coefficient, the higher the sensitivity of the system to the rate of change of the deviation, which helps to predict and suppress the change of the deviation in advance, but too large differential coefficient will lead to sensitivity to noise. In the injection process of the die casting machine, by reasonably setting and adjusting these parameters, precise control of the injection system can be achieved, and the production effect can be optimized.

[0073] S102, dividing the injection process into multiple injection stages according to the working state, after training the first model and the second model according to the historical process parameters, historical equipment parameters and historical control parameters of each injection stage, merging to obtain a third model;

[0074] It needs to be explained that the working state of the die casting machine such as injection speed, injection pressure, etc. changes with time, and its control parameters also change accordingly. There are great differences between the process parameters, equipment parameters and control parameters in different stages. By dividing the injection process of the die casting machine according to its working state, and analyzing and processing the parameters of each stage, the control parameters of the PID controller can be more accurately predicted.

[0075] S103, inputting the obtained real-time process parameters into the third model, and adjusting and optimizing the third model according to the real-time equipment parameters to obtain target control parameters, and adjusting the real-time control parameters of the injection system and / or switching the injection stage according to the target control parameters.

[0076] It should be noted that in the present embodiment, by arranging the injection displacement sensor, the injection pressure sensor, the plug-in valve position sensor, the mold temperature sensor, the mold cavity pressure sensor and the like arranged at the key positions of the die casting machine, the key process parameters and equipment parameters related to the injection process are collected in real time, the equipment parameter change trend in the future time period is obtained according to the real-time process parameters, the real-time equipment parameters are fed back to the first model, the output equipment parameter change trend of the first model is adjusted and optimized, and the target control parameter is obtained according to the equipment parameter change trend in the future time period after optimization.

[0077] In the present embodiment, the LSTM model and the RBF model are constructed to predict the control parameters of the PID controller, then the injection process is divided into multiple injection stages according to the working state of the injection system, the first model and the second model are iteratively trained according to the historical process parameters, equipment parameters and control parameters of each injection stage, the third model for predicting the control parameters is obtained, the model is trained in stages according to the parameter characteristics of different injection stages, so that the model can accurately predict the parameters of each stage, finally, the third model trained according to the real-time process parameters is used to adaptively adjust and optimize the output results of the first model in the third model according to the real-time equipment parameters monitored in real time, and finally, the target control parameter is obtained, the real-time control parameter of the PID is adjusted according to the target control parameter, and the working stage of the injection system is switched according to the adjusted real-time control parameter. In the present embodiment, an LSTM-RBF fusion model is constructed, the injection process of the injection system is divided into different injection stages, the control parameters of the PID controller are predicted according to the real-time process parameters, and the working stage of the injection system is switched according to the control parameters, so that the control parameters of the PID controller are accurately predicted and the control parameters and injection stages of the injection system are adaptively adjusted, and the accuracy of the injection system control is improved.

[0078] In some embodiments of the present application, the first model is constructed with the time sequence process parameters as input and the time sequence equipment parameters as output, including:

[0079] The first model is constructed based on the LSTM neural network, with the time sequence process parameters as input and the change trend of the equipment parameters in the future time period as output, and the first initial prediction model includes a first input layer, multiple LSTM layers, a full connection layer and a first output layer.

[0080] The input layer is used to receive historical time sequence process parameters.

[0081] It should be noted that the received time sequence process parameters are preprocessed, including but not limited to wavelet denoising technology to eliminate noise interference and improve data quality.

[0082] The LSTM layer is used to process the historical time-series process parameters, determine the time-series dependency in the time-series process parameters, and extract time-series features according to the dependency;

[0083] It should be noted that the LSTM layer is a core part of the LSTM neural network, which is composed of multiple LSTM units, each of which includes four key components: a forgetting gate, an input gate, a candidate memory unit, and an output gate, and the capture ability of complex time-series dependency is enhanced through a multi-layer cascading structure (such as 2-3 layers of LSTM).

[0084] The full connection layer is used to connect the LSTM layer and the output layer, and to perform feature mapping on the time-series features to obtain mapped features.

[0085] The output layer is used to output the change trend of the device parameters in the future time period according to the mapped features.

[0086] It should be noted that the first model analyzes and learns the relationship between the historical time-series process parameters and the historical device parameters, and predicts the change trend of the device parameters in the future time period according to the real-time process parameters.

[0087] In this embodiment, the LSTM neural network is used to construct the model, which can effectively save important features in the time series, analyze and learn process parameters and device parameters at different time points, and predict the change trend of device parameters in the future time period. This provides an accurate basis for subsequent data analysis.

[0088] In some embodiments of the present application, the first model is constructed by taking time-series process parameters as input and time-series device parameters as output, including:

[0089] The second model is constructed by taking device parameters as input and control parameters as output, including:

[0090] The second model is constructed based on the RBF neural network, taking the change trend of the device parameters as input and the control parameters as output, and the second initial prediction model includes a second input layer, multiple hidden layers, and a second output layer.

[0091] The second input layer is used to receive the change trend of the device parameters.

[0092] Specifically, the output result of the first model, the change trend of the device parameters in the future time period, is extracted through sliding window statistics (such as mean, variance, gradient) or time-series analysis (such as difference, moving average, etc.), and the dynamic features of the change trend of the device parameters are extracted, and the dynamic features and the real-time collected device parameters are normalized.

[0093] The hidden layer is used to extract feature information of the change trend of the device parameters.

[0094] The second output layer is used to output the control parameters according to the feature information.

[0095] It should be noted that the second output layer weights and sums the outputs of the hidden layers to obtain the control parameters of the PID controller, i.e., the proportional coefficient, the integral coefficient and the differential coefficient.

[0096] Specifically, the calculation formula of the second output layer is:

[0097]

[0098] wherein, is the weight of the i-th hidden layer to the output, is a bias term, and N is the number of hidden layers.

[0099] The embodiment analyzes and predicts the device parameter change trend in the future time period through the RBF neural network, determines the control parameters of the PID controller, and improves the accuracy of the control parameter prediction.

[0100] In some embodiments of the present application, as shown in Figure 2 , the present application provides an embodiment of the flowchart of step S102 in the method. Figure 2 Figure 1 The embodiment of the flowchart of step S102 in the method comprises:

[0101] S201, acquire historical time sequence process parameters and historical time sequence device parameters in different injection stages;

[0102] It should be noted that various sensors arranged at key positions of the die casting machine are used to collect the process parameters and device parameters in the injection process.

[0103] S202, iteratively train the first model according to the historical time sequence process parameters and the historical time sequence device parameters to obtain a first target model;

[0104] The historical time sequence process parameters are denoised and normalized, the number of layers and the number of units in each layer are set to optimize the learning ability and prediction accuracy of the network, then the LSTM neural network learns and captures the time sequence dependence in the data through the training of these historical data, and predicts the device parameter change trend in the injection process at the future time, i.e., the displacement amount of the servo plug-in valve core of the pressure accumulator, the displacement amount of the servo plug-in valve core of the booster accumulator and the pressure maintaining time requirement in the future time step.

[0105] S203, acquire historical device parameters and historical control parameters in different injection stages;

[0106] It should be noted that various sensors arranged at key positions of the die casting machine are used to collect the device parameters in the injection process, extract the control data in the historical work of the die casting machine, and use these historical data as training data.​​​

[0107] S204, iteratively training the second model according to the historical equipment parameters and the historical control parameters to obtain a second target model.

[0108] It should be noted that the RBF network is trained and learned through the die casting machine historical data, and the data input in the training and learning includes historical key equipment parameters (such as the injection accumulator servo plug-in valve core displacement amount, the booster accumulator servo plug-in valve core displacement amount, and the pressure maintaining time) and historical PID controller control parameters.

[0109] S205, the first target model and the second target model constitute a third model.

[0110] It should be noted that the trained LSTM first target model and the RBF second target model are combined to obtain an LSTM-RBF third model which predicts the control parameters, and at the same time, the third model is continuously optimized during the long-term operation of the injection system, and the LSTM and RBF network models are periodically updated by using the iterative learning strategy. The LSTM and the RBF network update the model by expanding the historical data and re-optimizing the network weight; at the same time, according to the actual running state of the system, the hyperparameters (such as the learning rate, the number of hidden layers, etc.) of the LSTM and the RBF neural network are further optimized to improve the adaptability and control precision of the system, and the frequency of the iterative learning is set according to the actual production demand, and the model is updated every certain period of time or when a large fluctuation of the injection process parameters is detected, so as to ensure the dynamic adaptive control ability of the system under complex working conditions.

[0111] In some embodiments of the present application, as shown in Figure 3 , the flowchart of an embodiment of step S202 in the present application is shown, which comprises: Figure 3 Figure 2

[0112] S301, based on the LSTM layer, forward propagating the historical time series process parameters to calculate the hidden state and the memory cell state of each time step;

[0113] Specifically, in the training process, the LSTM network will forward propagate the input die casting machine historical data (such as mold temperature, injection speed, and injection pressure), and calculate the hidden state and the memory cell state of each time step through the forward propagation of each unit of the LSTM layer.

[0114] S302, generating a predicted value through the full connection layer according to the hidden state and the memory cell state, and outputting the equipment parameter change trend of the future time period through the first output layer;

[0115] ​​S303, optimizing the first model according to a preset loss function, historical time sequence prediction device parameters and device parameter change trends in a future time period, to obtain a first optimized model;

[0116] Specifically, a mean square error loss function (MSE) is used to calculate the difference between the predicted value and the input historical key device parameters:

[0117]

[0118] wherein, is a loss value, N is the number of samples, are the true value and the predicted value of the t th sample, respectively.

[0119] S304, adjusting the output weight and bias of the first optimized model according to the gradient of the time step based on an optimization algorithm, to obtain a first target model.

[0120] Specifically, the gradient of each time step is calculated during the back propagation to each time step, and the gradient is calculated step by step forward propagation. The optimization algorithm is not specifically limited in the embodiment, and the Adam optimization algorithm is preferably used to update the weight and bias of the network. The calculation formula of the dynamic weight is:

[0121]

[0122] wherein, is the weight at t time, and are the momentum and the second moment estimation of the gradient, respectively, is a learning rate, is a small constant to prevent the denominator from being zero.

[0123] Further, in the iterative training process, if the loss function remains unchanged or becomes very small in multiple iterations, it is determined that the model has converged, and the training process is ended. The training effect is tested by using the learned weight and bias during training to perform forward propagation and generate a prediction result by inputting the time sequence data (such as mold temperature, injection speed and injection pressure) in the test set. If the error between the prediction result and the historical key device parameters in the test set is within 1%, it is determined that the training result is satisfactory. If the error is too large, the model parameters (such as the learning rate, the number of LSTM layers and the number of hidden layers) are adjusted or the training rounds are increased.

[0124] The performance of the model is optimized by iterative training of the LSTM model, and the accuracy of the control parameter prediction is improved.

[0125] In some embodiments of the present application, as shown in Figure 2 Figure 2 ​​​​The present invention provides Figure 1 The flowchart of an embodiment of step S204 includes:

[0126] S401, clustering historical device parameters based on a K-means clustering algorithm to obtain multiple cluster center points;

[0127] Specifically, multiple cluster centers are randomly initialized, the Euclidean distance between each device parameter and the center is calculated, the data is assigned to the nearest cluster, the cluster center is updated to the mean of the data within the cluster, and the algorithm is iterated until the center is stable to obtain multiple cluster centers.

[0128] S402, activating the hidden layer based on the Gaussian function and calculating the width parameter of the cluster center point;

[0129] Specifically, based on the Gaussian function, the width parameter of each center point is calculated as follows:

[0130] ,

[0131] in, is the input vector, It is The center vector of the cluster center points, It is The width parameter of the cluster center points, is a Gaussian function.

[0132] S403, training the weight of the second output layer according to the width parameter, and outputting the control parameter prediction result according to the weight;

[0133] Specifically, the output layer maps the output of the hidden layer to the target control parameters, namely the proportional, integral, and differential parameters of the PID controller; the output layer trains the weights by minimizing the objective function, which is formulated as follows:

[0134] ,

[0135] Among them, c is the control parameter, is the target output vector, i.e. the PID controller control parameters, is the weight matrix of the output layer, is the hidden layer output matrix, and this process is solved by the least squares method.

[0136] S404: Based on the preset loss function and control parameter prediction results, iteratively train the second model to obtain a second target model.

[0137] Specifically, the mean square error loss function is used to minimize the error between the output layer prediction value and the true value:

[0138] ,

[0139] wherein, N is the number of samples, and are the true output and the model prediction value of the i-th sample respectively, when the loss function value does not decrease significantly in continuous multiple training rounds, it is considered that the model converges, and the training process ends; the training effect is tested by a test set, the key equipment parameters (such as the displacement amount of the injection accumulator servo plug-in valve core, the displacement amount of the booster accumulator servo plug-in valve core, and the pressure maintaining time) in the test set are input, the prediction result is generated by using the weight learned during training; if the error between the prediction result and the PID controller control parameter in the test set is within 1%, it is determined that the training result is satisfactory, and if the error is too large, the model parameters (the center vector of the RBF neuron , the width parameter of the RBF neuron , the number of hidden layer neurons, etc.) are adjusted.

[0140] In some embodiments of the present application, the injection stage includes a slow injection stage, a fast injection stage, and a pressure boosting and pressure maintaining stage; the injection stage of the injection system is controlled according to the real-time control parameter, including:

[0141] When the real-time control parameter is in the preset slow injection interval, the injection system is controlled to be in the slow injection stage;

[0142] When the real-time control parameter is in the preset fast injection interval, the injection system is controlled to be in the fast injection stage;

[0143] When the real-time control parameter is in the preset pressure boosting and pressure maintaining interval, the injection system is controlled to be in the pressure boosting and pressure maintaining stage;

[0144] When the real-time control parameter reaches the preset pressure boosting and pressure maintaining upper limit, the injection system ends work.

[0145] It should be noted that in the injection process of the die casting machine, in the initial stage, the injection speed is low to avoid the splashing of the metal liquid and reduce the entrainment of gas, then the metal liquid is filled at high speed to ensure the completeness of the molding and reduce the residual gas in the cavity, and finally the pressure is rapidly boosted after the cavity is filled, the metal liquid is solidified under high pressure, the density of the casting is improved, until the pressure reaches the preset threshold, the pressure maintaining time reaches the preset threshold, and the injection process ends. Therefore, the injection process of the die casting machine is divided into a slow injection stage, a fast injection stage and a pressure boosting and pressure maintaining stage.

[0146] Specifically, according to the PID controller control parameter output by the third model, the position of the injection ram is controlled, as shown in Figure 5 . Figure 5The injection system control method provided by the present application comprises the following steps: when the position of the injection punch is in a preset slow injection interval, ensuring that the injection system is in a slow injection stage; when the position of the injection punch reaches a preset fast injection interval, switching the injection system to a fast injection stage; when the position of the injection punch reaches a preset pressure-increasing and pressure-maintaining interval, switching the injection system to a pressure-increasing and pressure-maintaining stage; and when both the pressure and the pressure-maintaining time reach preset upper limit values, ending the injection work.

[0147] In order to better implement the injection system control method in the embodiments of the present application, on the basis of the injection system control method, as shown in Figure 6 The present application also provides an injection system control device 600, which comprises:

[0148] A model construction module 601 is configured to construct a first model with time sequence process parameters as input and future time period equipment parameter variation trend as output, and construct a second model with future time period equipment parameter variation trend as input and control parameter as output.

[0149] A model training module 602 is configured to divide the injection process into multiple injection stages according to working states, and combine a third model after training the first model and the second model according to historical process parameters, equipment parameters and control parameters of each injection stage.

[0150] An injection system control module 603 is configured to obtain target control parameters by adjusting and optimizing the third model according to real-time process parameters, and adjusting real-time control parameters and / or switching injection stages according to the target control parameters.

[0151] The injection system control device 600 provided by the above embodiments can implement the technical solutions described in the above injection system control method embodiments, and the principles of the specific implementation of the above modules or units can be referred to the corresponding content in the above injection system control method embodiments, which will not be described here.

[0152] Those skilled in the art can understand that all or part of the processes of the above embodiments can be completed by a computer program to instruct related hardware (such as a processor, a controller, etc.) to complete, and the computer program can be stored in a computer readable storage medium. The computer readable storage medium includes a magnetic disk, an optical disk, a read-only memory or a random access memory, etc.

[0153] The injection system control method and device provided by the present application are described in detail above, and the principles and implementation manners of the present application are described by using specific examples. The above description of the examples is only used to help understand the method of the present application and the core idea thereof. Meanwhile, for those skilled in the art, the specific implementation manners and application ranges can be changed according to the idea of the present application. In summary, the content of the present description should not be understood as a limitation on the present application.

Claims

1. A control method for a shot system, characterized by, The method comprises the following steps: a first model is constructed by taking time sequence process parameters as input and future time period equipment parameter change trend as output, and a second model is constructed by taking future time period equipment parameter change trend as input and control parameter as output; a plurality of injection stages are divided according to working states, and the first model and the second model are trained according to historical process parameters, historical equipment parameters and historical control parameters of each injection stage, and then a third model is obtained by merging; real-time process parameters are input into the third model, the third model is adjusted and optimized according to real-time equipment parameters, target control parameters are obtained, and real-time control parameters of the injection system are adjusted and / or the injection stage is switched according to the target control parameters; a first model is constructed by taking time sequence process parameters as input and future time period equipment parameter change trend as output, which comprises: a first model is constructed by taking time sequence process parameters as input and future time period equipment parameter change trend as output based on an LSTM neural network, the first model comprises a first input layer, a plurality of LSTM layers, a full connection layer and a first output layer; the input layer is used for receiving historical time sequence process parameters; the LSTM layer is used for processing the historical time sequence process parameters, determining the dependence relationship of the time sequence in the time sequence process parameters, and extracting time sequence features according to the dependence relationship; the full connection layer is used for connecting the LSTM layer and the output layer, performing feature mapping on the time sequence features, and obtaining mapping features; the output layer is used for outputting the change trend of the equipment parameters in the future time period according to the mapping features; a second model is constructed by taking the change trend of the equipment parameters as input and the control parameter as output, which comprises: a second model is constructed by taking the change trend of the equipment parameters as input and the control parameter as output based on an RBF neural network, the second model comprises a second input layer, a plurality of hidden layers and a second output layer; the second input layer is used for receiving the change trend of the equipment parameters; the hidden layer is used for extracting feature information of the change trend of the equipment parameters; the second output layer is used for outputting the control parameter according to the feature information; the first model and the second model are trained according to historical process parameters, equipment parameters and control parameters of each injection stage, and then a third model is obtained by merging, which comprises: historical time sequence process parameters and historical time sequence equipment parameters of different injection stages are obtained; the first model is iteratively trained according to the historical time sequence process parameters and the historical time sequence equipment parameters, and a first target model is obtained; historical equipment parameters and historical control parameters of different injection stages are obtained; the second model is iteratively trained according to the historical equipment parameters and the historical control parameters, and a second target model is obtained; the first target model and the second target model constitute the third model; the first target model is obtained by iteratively training the first model according to the historical time sequence process parameters and the historical time sequence equipment parameters, which comprises: the historical time sequence process parameters are forward propagated based on the LSTM layer, and the hidden state and the memory cell state of each time step are calculated. The prediction value is generated by a full connection layer according to the hidden state and the memory cell state, and a future time period device parameter change trend is output by a first output layer; The first model is optimized according to a preset loss function, historical time sequence prediction device parameters and the future time period device parameter change trend, to obtain a first optimized model; The output weight and bias of the first optimized model are adjusted according to the gradient of the time step based on an optimization algorithm, to obtain a first target model; The optimization algorithm is an Adam optimization algorithm, and a calculation formula of the output weight is: ， where, is the weight at time t, and are the momentum and second moment estimate of the gradient, respectively, is the learning rate, is a small constant to prevent the denominator from being zero; The second model is iteratively trained according to the historical device parameters and the historical control parameters, to obtain a second target model, including: The historical device parameters are clustered based on a K-means clustering algorithm, to obtain a plurality of cluster center points; The hidden layer is activated based on a Gaussian function, and the width parameters of the cluster center points are calculated; The weight of the second output layer is trained according to the width parameters, and a control parameter prediction result is output according to the weight; The second model is iteratively trained based on a preset loss function and the control parameter prediction result, to obtain the second target model.

2. The injection system control method of claim 1 wherein, The third model is adjusted and optimized according to real-time device parameters, to obtain a target control parameter, including: Real-time device parameters in a pressure injection process of a die casting machine are acquired; The output result of the first model in the third model is adaptively optimized and adjusted according to the real-time device parameters; The output result of the first model after optimization and adjustment is taken as input of the second model, to obtain a final target control parameter.

3. The injection system control method of claim 1, wherein The pressure injection stage includes a slow pressure injection stage, a fast pressure injection stage and a pressure boosting and pressure maintaining stage, and the real-time control parameter is used to control the pressure injection stage of the pressure injection system, including: When the real-time control parameter is in a preset slow pressure injection interval, the pressure injection system is controlled to be in the slow pressure injection stage; When the real-time control parameter is in a preset fast pressure injection interval, the pressure injection system is controlled to be in the fast pressure injection stage; When the real-time control parameter is in a preset pressure boosting and pressure maintaining interval, the pressure injection system is controlled to be in the pressure boosting and pressure maintaining stage; When the real-time control parameter reaches a preset pressure boosting and pressure maintaining upper limit, the pressure injection system ends work.

4. An injection system control device for implementing the injection system control method according to any one of claims 1 to 3, characterized in that including: A model construction module is configured to construct a first model with time sequence process parameters as input and future time period device parameter change trend as output, and construct a second model with future time period device parameter change trend as input and control parameter as output; A model training module is configured to divide the pressure injection process into a plurality of pressure injection stages according to working states, and train the first model and the second model according to historical process parameters, device parameters and control parameters of each pressure injection stage, and then combine to obtain a third model; A pressure injection system control module is configured to obtain a target control parameter by adjusting and optimizing the third model according to real-time process parameters and real-time device parameters, and adjust the real-time control parameter and / or switch the pressure injection stage according to the target control parameter.

Images