阅读说明：本技术 一种压铸件缺陷预测及诊断系统 (Die casting defect prediction and diagnosis system ) 是由 程鹏 徐建华 叶盛 陈列 龙孟威 虞科 于 2021-06-04 设计创作，主要内容包括：本发明公开一种压铸件缺陷预测及诊断系统,包括一个或多个处理器获取工艺参数,通过深度学习模型对工艺参数进行处理与分析,输出预测的缺陷信息。所述深度学习模型还包括输入处理组件,用以对工艺参数进行加权处理和归一化处理。在上述预测及诊断系统的构建过程中还需要通过相关性分析对工艺参数进行筛选,以提高计算效率。本发明所涉及的方法考虑了多个压铸过程中的工艺参数,采用机器学习算法构建起工艺参数同缺陷之间的关联,可以预测压铸件的缺陷类型,同时能够实现在线检测工艺参数,并预测这些参数是否会在压铸件中产生缺陷,对于改进生产、实时监控产品状态、提高压铸件检测效率具有很好的工业价值,而且可以显著提高压铸行业的生产效率。(The invention discloses a die casting defect prediction and diagnosis system which comprises one or more processors, wherein the one or more processors are used for acquiring process parameters, processing and analyzing the process parameters through a deep learning model and outputting predicted defect information. The deep learning model further comprises an input processing component used for carrying out weighting processing and normalization processing on the process parameters. In the construction process of the prediction and diagnosis system, the process parameters need to be screened through correlation analysis so as to improve the calculation efficiency. The method provided by the invention considers the process parameters in a plurality of die-casting processes, adopts a machine learning algorithm to construct the correlation between the process parameters and the defects, can predict the defect types of the die-casting parts, can realize on-line detection of the process parameters, and can predict whether the parameters can generate defects in the die-casting parts, thereby having good industrial value for improving production, monitoring the product state in real time and improving the die-casting part detection efficiency, and remarkably improving the production efficiency of the die-casting industry.)

1. A die casting defect prediction and diagnosis system is characterized by comprising one or more processors, a defect detection module and a defect diagnosis module, wherein the one or more processors are used for acquiring process parameters; the one or more processors are to execute a deep learning model; the one or more processors inputting the process parameters into the deep learning model; the one or more processors executing the deep learning model predict defect information using the deep learning model; the one or more processors output the predicted defect information.

2. The die casting defect prediction and diagnosis system of claim 1, wherein the deep learning model comprises an input layer, an intermediate layer and an output layer, and the intermediate layer comprises a convolution layer, a pooling layer and a full connection layer.

3. The system for predicting and diagnosing the defects of the die casting as claimed in claim 1, wherein the process parameters are plural, and comprise the process parameters and the variation values of the corresponding parameters in the die casting process.

4. The die casting defect prediction and diagnosis system of claim 2, wherein the deep learning model further comprises an input processing component, and the input processing component is used for inputting the process parameters input by the one or more processors into an input layer of the deep learning model after data processing.

5. The die casting defect prediction and diagnosis system of claim 4, wherein said data processing comprises a weighting process, wherein said weighting process is a process parameter multiplied by a weighting factor for each process parameter.

6. The die casting defect prediction and diagnosis system according to claim 5, wherein the weight coefficient calculation steps of the process parameters are as follows:

step 1, collecting a data set, wherein the data set comprises technological parameters and corresponding die casting defect data;

step 2, pressing casting defectsYProcess parametersConstructing a linear model(ii) a Estimating by linear regressionIs given a value ofWeight coefficient of each process parameterRThe calculation method of (2) is as follows:

。

7. the die casting defect prediction and diagnosis system of claim 4, wherein said data processing comprises normalization processing.

8. The die casting defect prediction and diagnosis system of claim 7, wherein said normalization process employs a min-max normalization method.

9. The die casting defect prediction and diagnosis system of claim 3, wherein the process parameters comprise average low speed, average high speed, high speed starting point, material shank thickness, pressure build-up time, fill time, casting pressure, squeeze pin travel, mold clamping force, metal liquid temperature, system oil temperature, spray dose.

10. The die casting defect prediction and diagnosis system of claim 1, wherein the one or more processors acquire process parameters in real time through a process parameter acquisition module; the process parameter acquisition module comprises a data acquisition terminal and a data processing module unit; the data acquisition terminal acquires a plurality of process parameters and transmits the process parameters to the data processing module unit, and the data processing module unit processes the process parameters acquired by the data acquisition terminal and transmits the processed process parameters to the one or more processors; the data acquisition terminal acquires process parameters through communication of a plurality of sensors and/or a controller of a die casting machine or die casting machine matching equipment, wherein the communication comprises wired communication and wireless communication; the sensor is a combination of any one of a pressure sensor, pulse encoder data, a temperature sensor, a mold locking force sensor, a thermistor and an ultrasonic measuring instrument; and the data processing module unit processes the information acquired by the data acquisition terminal by adopting a Butterworth filtering algorithm.

11. The die casting defect prediction and diagnosis system of claim 1, wherein said defect information comprises an expected defect type.

12. The die casting defect prediction and diagnosis system of claim 11, wherein said expected defect types include no defect, die casting defect type, whether the die casting is a die casting in a hot die state; the die casting defect types comprise die casting surface defects and die casting internal defects; the surface defects of the die casting include but are not limited to cold pressing, undercasting, bubbles, depressions, flashes, layering, drawing marks, cracking, corner defects and burrs; the internal defects of the die casting include but are not limited to air holes, shrinkage cavities and slag inclusions.

13. The die casting defect prediction and diagnosis system of claim 12, wherein the expected defect types output by the deep learning model comprise a plurality of classifications, each including at least one defect type.

14. The die casting defect prediction and diagnosis system of claim 11, wherein said one or more processors perform said defect cause rule search based on said predicted defect information; searching and outputting the reason causing the expected defect type in a die casting quality database according to the predicted defect information by the defect reason rule search; the cause causing the desired defect type is one cause or a plurality of causes or a ranking of the causes; the reasons for the expected defect type comprise any one or more of items, deviation directions and suggested value ranges of process parameters causing the die casting to have the expected defect type.

15. A method for screening process parameters according to claim 3, comprising the steps of:

step 1, collecting technological parameters in a die-casting process;

and 2, identifying the process parameters with strong correlation through the mechanical structure and the physical law of the die casting machine or analyzing the correlation among the process parameters through a grey correlation degree analysis method to identify the process parameters with strong correlation, and then only reserving one of the process parameters for two or more process parameters with strong correlation.

16. A screening method according to claim 15, further comprising the steps of:

step 3, collecting die casting defect data corresponding to the plurality of process parameters obtained by screening in the step 2, and carrying out correlation analysis on the die casting defect condition and the plurality of process parameters; the correlation analysis method is multivariate correlation analysis, and identifies process parameters with strong correlation with the defect condition of the die casting as the process parameters input into the deep learning model.

17. The screening method of claim 16, wherein said strong correlation is a correlation coefficient greater than 0.75.

18. A method of training a die casting defect prediction and diagnosis system according to any one of claims 1 to 13, comprising the steps of:

step 1, constructing a training sample, wherein the training sample comprises technological parameters and corresponding defect data of a die-casting product;

step 2, inputting the process parameters into a deep learning model to obtain defect information output by the deep learning model;

step 3, calculating the prediction accuracy of the defect information output by the deep learning model through a mean square error function;

step 4, optimizing the deep learning model by adopting any one or combination of random gradient descent method, BPTT algorithm, forward propagation algorithm and backward propagation algorithm, and reducing the error between the deep learning model and the defect sample by optimizing the weight and bias of the deep learning model; and when the error is stable, finishing the training of the deep learning model for predicting the die-casting defect training.

19. The training method of claim 18, wherein the mean square error function in step 3 is used to calculate the error between the defect information output by the deep learning model and the defect data of the medium-pressure casting in the training sample, so as to obtain the prediction accuracy of the deep learning model.

20. A training method as claimed in claim 18, wherein the mean square error function in step 3 is used to calculate the error between the defect information output by the deep learning model, so as to obtain the prediction accuracy of the deep learning model.

Technical Field

The invention relates to the field of die casting, in particular to a die casting defect prediction and diagnosis system.

Background

The die casting process has many process parameters which need to be controlled, and the process parameters are related to each other and are a complex whole, and when a certain parameter is changed, the quality problem of die castings can be caused. With the continuous promotion of customer to product quality requirement, it is more and more important to the die casting machine to ensure that product quality is stable.

At present, in actual use, the machined die casting is mainly inspected and analyzed in a sampling inspection mode, the process is repeatedly adjusted and optimized according to an inspection result, and a large amount of detection time needs to be consumed. Moreover, the detection method does not relate to the analysis of various process parameters in the production process, cannot establish the relationship between the defects of the die casting and the process parameters, and cannot predict the quality condition of the die casting according to the process parameters. The sampling inspection mode also has the problems of multiple sampling inspection, low efficiency and the like, and meanwhile, the online process quality inspection cannot be carried out.

Control charts are also a common tool in the field of quality control. The control map is a map with control limits used for analyzing and judging whether the process is in a steady state, and is a functional chart having a function of distinguishing normal fluctuations from abnormal fluctuations. The control chart can be used for controlling the change condition of the quality characteristic value in the production process to see whether the process is in a stable and controlled state; and whether an abnormal condition occurs in the production process can be found so as to prevent unqualified products from being produced. However, the traditional control diagram method is difficult to use in actual production due to the fact that the die-casting environment is relatively severe, the operation is complex, and the automation level of the die-casting machine is limited.

Disclosure of Invention

The invention aims to provide a die casting defect prediction and diagnosis system to solve the problems in the prior art. In order to achieve the purpose, the specific technical scheme of the invention is as follows:

a die casting defect prediction and diagnosis system comprises one or more processors, a defect detection module and a defect diagnosis module, wherein the one or more processors are used for acquiring process parameters; the one or more processors are to execute a deep learning model; the one or more processors inputting the process parameters into the deep learning model; the one or more processors executing the deep learning model predict defect information using the deep learning model; the one or more processors output the predicted defect information.

Preferably, the deep learning model comprises an input layer, an intermediate layer and an output layer, wherein the intermediate layer comprises a convolutional layer, a pooling layer and a full-link layer.

Preferably, the process parameter is a plurality of parameters, including the process parameter and the variation value of the corresponding parameter in the die casting process.

Preferably, the deep learning model further comprises an input processing component, and the input processing component is used for inputting the process parameters input by the one or more processors into an input layer of the deep learning model after data processing.

Preferably, the data processing includes a weighting process of multiplying the process parameters by a weight coefficient of each process parameter.

Preferably, the step of calculating the weight coefficient of the process parameter is as follows:

step 1, collecting a data set, wherein the data set comprises technological parameters and corresponding die casting defect data;

step 2, pressing casting defectsYProcess parametersConstructing a linear model(ii) a Estimating by linear regressionIs given a value of. Weight coefficient of each process parameterRThe calculation method of (2) is as follows:

。

preferably, the data processing comprises normalization processing.

Preferably, the normalization process employsmin-maxA method of standardization.

Preferably, the process parameters include average low speed, average high speed, high speed starting point, material shank thickness, pressure build-up time, filling time, casting pressure, squeeze pin travel, mold clamping force, metal liquid temperature, system oil temperature, spray dose.

Preferably, the one or more processors acquire the process parameters in real time through the process parameter acquisition module; the process parameter acquisition module comprises a data acquisition terminal and a data processing module unit; the data acquisition terminal acquires a plurality of process parameters and transmits the process parameters to the data processing module unit, and the data processing module unit processes the process parameters acquired by the data acquisition terminal and transmits the processed process parameters to the one or more processors; the data acquisition terminal acquires process parameters through communication of a plurality of sensors and/or a controller of a die casting machine or die casting machine matching equipment, wherein the communication comprises wired communication and wireless communication; the sensor is a combination of any one of a pressure sensor, pulse encoder data, a temperature sensor, a mold locking force sensor, a thermistor and an ultrasonic measuring instrument; and the data processing module unit processes the information acquired by the data acquisition terminal by adopting a Butterworth filtering algorithm.

Preferably, the defect information includes a desired defect type.

Preferably, the expected defect type comprises no defect, die casting defect type, whether the die casting is in a hot die state or not; the die casting defect types comprise die casting surface defects and die casting internal defects; the surface defects of the die casting include but are not limited to cold pressing, undercasting, bubbles, depressions, flashes, layering, drawing marks, cracking, corner defects and burrs; the internal defects of the die casting include but are not limited to air holes, shrinkage cavities and slag inclusions.

Preferably, the expected defect type output by the deep learning model comprises a plurality of classifications, and each classification comprises at least one defect type.

Preferably, the one or more processors perform the defect cause rule search according to the predicted defect information; searching and outputting the reason causing the expected defect type in a die casting quality database according to the predicted defect information by the defect reason rule search; the cause causing the desired defect type is one cause or a plurality of causes or a ranking of the causes; the reasons for the expected defect type comprise any one or more of items, deviation directions and suggested value ranges of process parameters causing the die casting to have the expected defect type.

A screening method of process parameters for a die casting defect prediction and diagnosis system comprises the following steps:

step 1, collecting technological parameters in a die-casting process;

and 2, identifying the process parameters with strong correlation through the mechanical structure and the physical law of the die casting machine or analyzing the correlation among the process parameters through a grey correlation degree analysis method to identify the process parameters with strong correlation, and then only reserving one of the process parameters for two or more process parameters with strong correlation.

Preferably, the method further comprises a step 3 of collecting die casting defect data corresponding to the plurality of process parameters obtained by screening in the step 2, and performing correlation analysis on the die casting defect condition and the plurality of process parameters; the correlation analysis method is multivariate correlation analysis, and identifies process parameters with strong correlation with the defect condition of the die casting as the process parameters input into the deep learning model.

Preferably, the strong correlation means that the correlation coefficient is greater than 0.75.

A training method of a die casting defect prediction and diagnosis system comprises the following steps:

step 1, constructing a training sample, wherein the training sample comprises technological parameters and corresponding defect data of a die-casting product;

step 2, inputting the process parameters into a deep learning model to obtain defect information output by the deep learning model;

step 3, calculating the prediction accuracy of the defect information output by the deep learning model through a mean square error function;

step 4, optimizing the deep learning model by adopting any one or combination of random gradient descent method, BPTT algorithm, forward propagation algorithm and backward propagation algorithm, and reducing the error between the deep learning model and the defect sample by optimizing the weight and bias of the deep learning model; and when the error is stable, finishing the training of the deep learning model for predicting the die-casting defect training.

Preferably, the mean square error function in the step 3 is used for calculating an error between defect information output by the deep learning model and defect data of the medium-pressure casting in the training sample, so as to obtain the prediction accuracy of the deep learning model.

Preferably, the mean square error function in step 3 is used to calculate an error between defect information output by the deep learning model, so as to obtain the prediction accuracy of the deep learning model.

Compared with the prior art, the die casting defect prediction and diagnosis system takes the process parameters in a plurality of die casting processes into consideration, and adopts a machine learning algorithm (Deep Learning Algorithm) And establishing a correlation between the process parameters and the defects, predicting whether the die casting is defective or not, and expecting the defect type. Meanwhile, the method can realize on-line detection of the process parameters and predict whether the parameters can generate defects in the die casting, has good industrial value for improving production, monitoring the product state in real time and improving the die casting detection efficiency, and can show that the method can be used for detecting the process parameters on lineThe production efficiency of the die-casting industry is improved.

In order to make the present invention comprehensible, preferred embodiments accompanied with figures are described in detail below.

Drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings needed to be used in the description of the embodiments will be briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings based on these drawings without creative efforts.

FIG. 1 is a system schematic of one embodiment of the present invention;

FIG. 2 is a schematic diagram of a method for screening process parameters according to an embodiment of the present invention;

FIG. 3 is a schematic diagram of the construction, training and use process of a prognosis and diagnosis system according to yet another embodiment of the present invention;

FIG. 4 is a diagram illustrating the use of a deep learning model according to another embodiment of the present invention.

Detailed Description

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

As shown in fig. 1, the system for predicting and diagnosing the defects of the die castings comprises one or more processors for acquiring process parameters; the one or more processors are to execute a deep learning model; the one or more processors inputting the process parameters into the deep learning model; the one or more processors executing the deep learning model predict defect information using the deep learning model; the one or more processors output the predicted defect information.

The deep learning model comprises an input layer, a middle layer and an output layer, wherein the middle layer comprises a convolution layer, a pooling layer and a full-connection layer.

Inputting the process parameters into an input layer in the deep learning model to obtain process parameter vectors; inputting the process parameter vector into a convolutional layer in the deep learning model, so as to perform feature extraction on the process parameter vector based on at least one feature dimension and obtain the process parameter feature; inputting the process parameter characteristics into a pooling layer in the deep learning model to obtain defect characteristics through screening; inputting the defect features into a full-link layer in the deep learning model to determine the probability that the defect features belong to defect information categories; and then outputting the defect information through the output layer.

The process parameters are multiple, and include process parameters and variation values of corresponding parameters in the die-casting process, generally process parameters related to pressure, speed, time and temperature in the die-casting process.

In some implementations, data processing of the process parameters input into the deep learning model is required to achieve a balance between accuracy and efficiency of the computer-implemented method to which the present invention relates.

The deep learning model further comprises an input processing component, and the input processing component is used for inputting the process parameters input by the one or more processors into an input layer of the deep learning model after data processing.

The data processing includes weighting processing and normalization processing.

The weighting process is to multiply the process parameters by a weight coefficient of each process parameter.

And the normalization processing is to directly normalize the process parameters input into the deep learning model or the process parameters after the weighting processing.

The normalization process adoptsmin-maxNormalization method, concrete transfer function:

in the above formula, the first and second carbon atoms are,as specific values of the process parameters or the process parameters of the weighting process,is the minimum value in the parameter set of the process parameter or the minimum value after the parameter set of the process parameter is weighted,the maximum value in the parameter set of the process parameter or the maximum value after the parameter set of the process parameter is weighted,and normalizing the corresponding process parameters into dimensionless characteristic parameters between 0 and 1.

In some implementations, to achieve better prediction accuracy for computer-implemented methods of the present invention, the process parameters include average low speed, average high speed, high speed starting point, shank thickness, build time, fill time, casting pressure, squeeze pin travel, mold clamping force, metal liquid temperature, system oil temperature, spray dose. The average low speed is the average low speed of injection, the default is a speed interval with the average speed lower than 1m/s, and the average median value of data is calculated in a continuous time period during acquisition; the average high speed is the average high speed of injection, and the average high speed is calculated in the process speed interval which is reduced by 8% of the fluctuation of the maximum speed in the patent; the high-speed starting point is a position when the injection speed is switched from low speed to high speed, and the high-speed starting point in the patent uses a high-speed starting point calculated by 0.1ms of delay in the process of switching the low speed to the high speed; the thickness of the material handle refers to the length of the residual material in the pressure chamber after the molten metal is filled into the die cavity, and the difference value between the air compression injection stroke and the injection end position is used as a calculated value in the patent; the pressure build-up time refers to the time from the end of a rapid injection stage to the end of a pressurization stage of the die casting machine hammer head casting pressure, and is specifically represented as the time from the rise of the casting pressure from specified pressure to 90% of set pressure, and the time from the arrival of inlet pressure to a set value after the end of rapid injection, wherein in the calculation process, the time interval from the time when the position of the hammer head is less than 0.05mm to the time when the steady-state pressure reaches 90% is taken as a calculated value; the filling time is the time from the time when the molten metal reaches the inner pouring gate to the time when the molten metal fills the mold; the casting pressure refers to the pressure of the injection hammer head, namely the pressure of molten metal; the extrusion pin stroke refers to the stroke of a core-pulling module for feeding molten metal in the mold; the mold clamping force refers to the mold clamping force of a mold clamping mechanism in a high-pressure mold clamping state when a die casting machine is in a die-casting high-pressure state; the temperature of the molten metal refers to the temperature of molten aluminum in the process; the system oil temperature refers to a transmission medium for driving a mechanical structure in a die casting machine hydraulic system and is mainly expressed as the temperature of oil; the spraying dosage is the dosage of spraying a corresponding release agent on the surface of the mold, and mainly represents the spraying thickness covered by the surface of the mold, and the release agent is mainly organic liquid used for the surface of the mold.

In some implementations, the one or more processors obtain the process parameters in real time through a process parameter acquisition module; the process parameter acquisition module comprises a data acquisition terminal and a data processing module unit; the data processing module unit processes the technological parameters acquired by the data acquisition terminal and transmits the processed technological parameters to the one or more processors. And the data acquisition terminal acquires process parameters through communication of a plurality of sensors and/or a controller of the die casting machine or the die casting machine matching equipment. The communication includes wired communication and wireless communication, and because the inside data format that adopts of die casting machine or die casting machine corollary equipment, transfer protocol is different, data acquisition terminal need set up a plurality of connected modes and data protocol according to the use scene. The sensor is the combination of any one of a pressure sensor, pulse encoder data, a temperature sensor, a mold locking force sensor, a thermistor and an ultrasonic measuring instrument. And the data processing module unit processes the information acquired by the data acquisition terminal by adopting a Butterworth filtering algorithm.

The defect information includes a desired defect type. The expected defect types comprise no defect, die casting defect types and whether the die casting is in a hot die state.

The defect types of the die casting comprise pressure cold shut, undercasting, bubbles, depressions, flashes, layering, drawing marks, cracking, corner defects, burr isopiestic surface defects of the casting; internal defects of castings such as air holes, shrinkage cavities and slag inclusions.

The die casting has a die preheating stage in the production process, namely a hot die state, the die casting produced in the hot die state is a special quality defect, and identification and prediction are needed, wherein whether the die casting is in the hot die state or not is used for indicating whether the die casting is in the hot die state or not.

Because die casting defect types are more, in some implementation modes, in order to improve the training efficiency and the calculation efficiency of the computer implementation method, expected defect types output by the deep learning model are classified as follows: type 1 indicates die casting surface defects such as cold shut, undercast, bubbles, dimples, flashes, delamination, pull marks, cracks, corner defects, burrs; type 2 indicates internal defects of the die casting, such as blowholes, air holes, shrinkage cavities, slag inclusions; type 3 indicates that the die casting is a die casting in a hot die state of a die casting machine; type 4 indicates that the die cast product was good and free of defects.

In some implementations, the one or more processors may also perform the defect cause rule search based on the predicted defect information.

And searching and outputting the reasons causing the expected defect types in the die casting quality database according to the predicted defect information by the defect reason rule search, wherein the reasons include any one or more of items, deviation directions and suggested value ranges of process parameters causing the expected defect types of the die castings.

The die casting quality database is derived from industry and enterprise experiences and comprises a data set of defect types and corresponding reasons, for example, predicted defect information output by a deep learning model is 'air holes', the main reason for generating the air holes with poor industry experience is insufficient casting pressure, and 'defect-air holes and defect reasons' are stored in the die casting quality database: casting pressure "; the defect reason rule output by the system can be casting pressure, namely the defect occurs due to the setting of the casting pressure, can also be casting pressure set value lower than the process standard, namely the current casting pressure set value is lower than a normal value (value which does not cause the defect), and can also be casting pressure set value recommended to be 60-70 MPa, namely a recommended value range is provided. It should be noted that the main reason for the generation of the bad air holes is that the insufficient casting pressure is only an example, and the actual output result may include a plurality of reasons or a sequence of the reasons.

It should be noted that, the corresponding relationship between the die casting defect and the process parameter in the die casting quality database from the industry and enterprise experience is generally unidirectional, qualitative, that is, the main reason for the generation of the bad air hole according to the industry experience is the insufficient casting pressure, but the industry experience cannot establish the data association between the casting pressure and the bad air hole, that is, when the process parameter is set, whether the bad air hole appears or not cannot be predicted based on the setting of the casting pressure. The phenomenon widely occurs in various fields, such as the health field, when a certain organ of a human body has obvious pathological changes, such as fatty liver, the etiology, such as alcoholic fatty liver, can be relatively easily determined according to living habits and other examinations, but under the technical conditions of the current hospital, the accurate probability of the fatty liver occurring is difficult to predict for a healthy individual.

As the die-casting process has numerous process parameters, if the process parameters which can be collected are all input into the deep learning model, the training period is longer, the calculation efficiency is lower, and the requirement on hardware for operating the deep learning model is higher, so that the invention provides a method for screening the process parameters, as shown in fig. 2, which comprises the following steps:

step 1, collecting process parameters, and establishing a first parameter set (s 1, s2 … … … … si); the process parameters collected in one embodiment include the following 17: average low speed s1, average high speed s2, maximum speed s3, high speed start s4, shank thickness s5, pressure build-up time s6, filling time s7, casting pressure s8, inlet pressure s9, extrusion pin stroke s10, die temperature s11, spray amount s12, spray frequency s13, mold locking force s14, system oil temperature s15, metal liquid temperature s16, and injection stroke s 17.

And 2, analyzing the correlation of the process parameters, wherein the number of the process parameters which can be used in practice is often large, some process parameters among the process parameters have direct or indirect correlation, and part of the correlation can be calculated by the mechanical structure, the physical rule and the like of the die casting machine, such as the relation between the speed and the time. Therefore, correlation analysis needs to be performed on the collected process parameters to identify the process parameters with strong correlation. Such as average speed s2 and maximum speed s3, casting pressure s8 and inlet pressure s9 have a strong correlation and similar patterns. The correlation between the process parameters can also be analyzed by a correlation analysis method, such as a grey correlation analysis method, to identify the process parameters with strong correlation. Thereafter, for two or more process parameters having a strong correlation, only one of the process parameters is retained, and the second parameter sets, i.e. the average high speed s2 and the maximum speed s3, and only the average high speed s2 and the casting pressure s8 of the casting pressure s8 and the inlet pressure s9 are retained.

And 3, collecting die casting defect data corresponding to the process parameters in the second parameter set, wherein the data are from historical data of equipment, and the defect condition is obtained through manual detection. Then, carrying out correlation analysis on the defect condition of the die casting and a plurality of process parameters; the correlation analysis method is multivariate correlation analysis and identifies the process parameters with strong correlation with the defects, wherein the strong correlation means that a correlation coefficient R is more than 0.75. And selecting the process characteristic parameters with the correlation coefficient R larger than 0.75 to obtain a third parameter set. The process parameters in the third parameter set are the process parameters input to the deep learning network.

In some implementations, in order to achieve the balance between the accuracy and the efficiency of the computer-implemented method according to the present invention, a weighting process needs to be performed on the process parameters input to the deep learning network, that is, the process parameters in the third parameter set, where the weight coefficient of each process parameter is calculated as follows:

step 1, using die casting defect data (variables are recorded as)Y) Process parameters (variables are noted) The linear model was constructed as follows:；

step 2, estimating by adopting a linear regression modeThe value of (2) is obtained by estimating the auxiliary parameter by a least square method；

Step 3, weighting coefficient of each process parameterRThe calculation method of (2) is as follows:

cov represents covariance, var represents variance, and the correlation value of the die casting defects and the process characteristic parameters is determined by the formula. It should be noted that, the calculation process of the correlation coefficient R obtained by the multivariate correlation analysis calculation of the die casting defect condition and the plurality of process parameters mentioned above and the weight coefficient of each process parameterThe calculation process of R is the same.

The invention provides a method for training the predictive die-casting defect depth calculation model, which comprises the following steps:

step 1, constructing a training sample, wherein the training sample comprises process parameters and defect data of die castings produced by the process parameters, and the data is from historical data of equipment;

step 2, inputting the process parameters in the training sample into a deep learning model to obtain defect information output by the deep learning model;

step 3, calculating the prediction accuracy of the defect information output by the deep learning model through a mean square error function;

step 4, optimizing the deep learning model by adopting any one or combination of random gradient descent method, BPTT algorithm, forward propagation algorithm and backward propagation algorithm, and reducing the error between the deep learning model and the defect sample by optimizing the weight and bias of the deep learning model; and when the error is stable, finishing the training of the deep learning model for predicting the die-casting defect training.

In some implementations, the input layer data of the deep learning model is X_i = {x_1，x_2，……x₁₂The intermediate layer data(ii) a The output layer of the deep learning model is，For the expected defect types, for improved computational efficiency, the expected defect types are divided into four types, as described above, with data labels of。

The weight from the input layer to the middle layer is(ii) a The weight from the middle layer to the output layer is(ii) a A bias coefficient of。

The deep learning model mathematical model is as follows:

in the above model F₁For the activation function of the input layer to the intermediate layer, F₂Is the activation function of the middle layer to the output. F₁(x) 、F₂(x) GetSigmoid(x) The function of the function is that of the function,Sigmoidthe function is often used as an activation function for neural networks, mapping variables between 0 and 1.

G. M is an intermediate variable parameter, and M is an intermediate variable parameter,for loss functions, RMSE is the root mean square error.

The random gradient descent method is adopted on the gradient descent method and strategy, and the random gradient descent means that one sample is randomly drawn from all samples at a time to obtain a gradient, and the gradient is used for updating.

Wherein, W is a parameter to be updated;tis as followstPerforming secondary iteration;is as followstThe learning rate of the sub-iteration is,representing a gradient operator;is a randomly sampled sample loss function.

In some implementations, the BPTT algorithm model employed is as follows:

wherein、、In order to input the parameters of the deep learning model,is thatThe input of the time of day is,、、is the intermediate variable(s) of the variable,andthe BPTT algorithm for the activation functions input to the middle tier and output from the middle tier, respectively, is:

i.e. the updating of the parameters at a certain time t, requires the gradient at each previous time to participate in the operation.

In some implementations, the forward propagation algorithm performs a series of linear operations and activation operations using a plurality of weight coefficient matrices and bias vectors and input value vectors, starting from an input layer, calculating backward layer by layer, and operating until an output result is obtained. The back propagation algorithm is a process of carrying out iterative optimization on a loss function of the deep learning model by using a gradient descent method to obtain a minimum value.

In some implementations, the mean square error function in step 3 is used to calculate an error between the defect information output by the deep learning model and the defect data of the medium-pressure casting in the training sample, so as to obtain the prediction accuracy of the deep learning model.

In some implementations, the mean square error function in step 3 is used to calculate the error between the defect information output by the deep learning model to obtain the prediction accuracy of the deep learning model

First, it is clear that Root Mean Squared Error (RMSE) is the square Root of the Mean square Error, representing the deviation of the predicted valueThe degree of divergence, also known as the standard error, best fit case is RMSE = 0. The root mean square error is defined as follows (where E _i To representiAbsolute error of each reference value from the predicted value;Y _i to representiA reference value;to representiIndividual predicted values):

the construction, training (including the above process parameter screening and training methods) and use of the predictive diagnostic system according to the present invention are briefly described as follows with reference to the following examples:

before the deep learning model is built and trained, input process parameters need to be determined, relevant process parameters are collected through a process parameter collecting module in the deep learning model, an obtained data set comprises process parameter data, wherein the process parameters are Si=（S_1，S_2，……S_i) And the variation value of the pressure casting defect data, and the pressure casting defect data under the process parameters, wherein the pressure casting defect data is obtained by manually detecting the pressure casting, and comprises the defect-free data, the defect type of the pressure casting and whether the pressure casting is in a hot die state. For convenience of numerical calculation and statistics, the defect type of the die casting may be defined as a numerical value, such as "0" for no defect, "1" for defect type of die casting, and "2" for die casting in a hot die state or not.

The process parameters in the data set need to be screened by the screening method described above before training, so as to improve the calculation efficiency. Meanwhile, in order to balance the calculation efficiency and the calculation precision, the weight coefficient of the screened process parameters is obtained by adopting the weight coefficient calculation method. The weight coefficients are used to construct the input processing components in the deep learning model described above. And forming a training sample by using the screened process parameters and the corresponding die casting defect data thereof, wherein multiple groups of data are generally required to achieve the training effect, and after the training sample is obtained, the multiple groups of data are divided into a training group and a verification group.

For example, obtaining an initial parameter data set of S = (S)₁，S₂，S₃，S₄，S₅，S₆，S₇，S₈) Determination of S by correlation analysis of Process parameters_4、S₅The correlation between the two process parameters is high, and only S is reserved₄。

The die casting defect data corresponding to the data set technological parameters are the same as S₁，S₂，S₃，S₄，S₆，S₇，S₈Performing multivariate correlation analysis, wherein multivariate analysis includes multiple sets of data, i.e. data of multiple die casting processes, and collecting S in each die casting process₁，S₂，S₃，S₄，S₆，S₇，S₈And the data of the defects of the die castings produced by the secondary die casting process. Calculating to obtain a correlation coefficient R = (R)₁，R₂，R₃，R₄，R₆，R₇，R₈) = (0.8, 0.9, 0.2, 0.75, 0.2, 0.75, 0.8) according to R_i>0.75, i.e. the screened data S = (S)₁，S₂，S₄，S₇，S₈) On the basis, the process parameter in the training sample is determined to be W = (S)₁，S₂，S₄，S₇，S₈）。

And determining a weight coefficient in the input processing component according to the calculated correlation coefficient. And then, the deep learning model can be trained through the training samples, and in the training process, the weight does not need to be manually input when data is input into the deep learning model.

The normalization process input into the processing component is illustrated as follows, where the process parameters include process parameters and variation values of the corresponding parameters in the die casting process, and in practice, the process parameter acquisition module performs sampling according to a preset period and transmits the process parameters to the deep learning model according to the preset period.

For example, 5 consecutive sets of process parameters are entered as follows:

normalized data:

in the training of the deep learning model, the deep learning model is firstly trained by using data of a training set, and the output result of the deep learning model isY _outDuring the training processY _outIt may be any one or more of the above "0, 1, 2", or it may be the probability that the die casting defect type is the desired defect type, such as being defect free, for example 90% being defect free.

In one embodiment, a mean square error function is adopted to calculate the error between the defect information output by the deep learning model and the defect data of the medium-pressure casting in the training sample, so as to obtain the prediction accuracy of the deep learning model. And adjusting the deep learning model according to the accuracy.

In another embodiment, the mean square error function is used to calculate the error between the defect information output by the deep learning model, i.e. the defect information output by the deep learning model is for the same defect typeY _outError therebetween, e.g. data predicting cold trap defect type for a plurality of sets, whichY _outThe numerical value of (A) is not 0.961, 0.986, 0.996 and the like, and the numerical value of (A) is adjusted continuously to aim at the same defect typeY _outWithin a defined range, e.g. cold shutY _outHas a value of 0.9605-0.9609.

And when the error is stable, verifying the transmission result of the deep learning model by using the verification group data, if the verification result is qualified, for example, aiming at 200 verification groups of data, 199 times of outputs of the deep learning model are correct, namely, the training of the deep learning model for predicting the die-casting defect training is completed.

As shown in fig. 3 and 4, the trained deep learning model can be used for defect prediction, and when prediction is performed, the process parameter acquisition module acquires the process parameters on line in real time and transmits the process parameters to the deep learning model, wherein the process parameters are the process parameters which are determined by screening through the screening method. And outputting predicted defect information based on the process parameter deep learning model. Based on the defect information, the cause causing the desired defect type is searched and output through the defect cause rule described earlier.

The following table gives an illustrative example of the prediction, wherein the expected defect type is output by a deep learning model according to definition, the reason causing the expected defect type is that according to the experience of the current production process, relevant principles are induced for the defect content of the die casting, the principles are stored in a die casting quality database, and a processor searches a reason rule corresponding to the expected defect type in the die casting quality database according to the expected defect type and outputs the search result. It should be noted that the defect cause rules given in the following table are only examples, and the actual output result may include multiple causes or the sequence of multiple causes.

Table 1: predicted condition of output

Output of	Expected defect type	Cause of the expected defect type
			Value1	Casting surface defect-cold shut	Too low casting pressure
Value2	Casting surface defect-mark	Mean too fast
			Value3	Casting surface defect-undercasting	Low casting pressure
Value4	Casting surface defect-dishing	The extrusion pin has smaller stroke
			Value5	Casting surface defect-flash	The clamping force is smaller
Value6	Casting surface defect-delamination	Too low casting pressure
			Value7	Surface defect-cracking of castings	Too short spraying time
Value8	Surface defect-burr of casting	Mean high velocity too great
			Value9	Internal defects-pores of castings	Average low speed too fast
Value10	Internal defect-shrinkage cavity of casting	Excessive temperature of the molten metal
			Value11	Die casting in hot die state	Preheating die temperature, waste products

The present invention provides a computer-readable medium storing computer instructions for predicting die-casting defects, which when executed by one or more processors, cause the one or more processors to perform the above-described computer-implemented method for predicting die-casting defects.

Implementations of the subject matter and the functional operations described in this specification and the claims may be implemented in digital electronic circuitry, tangible computer software or firmware, computer hardware, or in combinations of one or more of them.

As used in the description and claims of the present application, the terms "processor," "data processing module," "computer," or "electronic computing device" (or equivalent terms as understood by those skilled in the art) refer to data processing hardware, including various devices, apparatus, and machines for processing data, including by way of example a programmable processor, a computer, or multiple processors or computers. The apparatus can also include, or further include, special purpose logic circuitry, e.g., a Central Processing Unit (CPU), an FPGA (field programmable gate array), or an ASIC (application-specific integrated circuit). In some implementations, the data processing apparatus or dedicated logic circuit (or a combination of data processing apparatus or dedicated logic circuit) may be hardware or software based (or a combination of hardware and software based). The apparatus can optionally include code that creates an execution environment for the computer program, e.g., code that constitutes processor firmware, a protocol stack, a database management system, an operating system, or a combination of execution environments. The present invention contemplates the use of a data processing device with or without a conventional operating system such as LINUX, UNIX, WINDOWS, MAC OS, ANDROID, IOS or any other suitable conventional operating system.

Implementations of the subject matter described in this specification can be implemented in a computing system that includes a back end component (e.g., as a data server), or that includes a middleware component (e.g., an application server), or that includes a front end component, e.g., a client computer having a graphical user interface or a web browser through which a user can interact with an implementation of the subject matter described in this specification, or any combination of one or more such back end, middleware, or front end components. The components of the system can be interconnected by any form or medium of wired or wireless digital data communication (or combination of data communication), e.g., a communication network.

The computing system may include clients and servers. A client and server are generally remote from each other, and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other.

While the specification and claims herein contain many specific implementation details, these should not be construed as limitations on the scope of any invention or of what may be claimed, but rather as descriptions of features specific to particular implementations of particular inventions. In the previously described implementations, the separation or integration of various system modules and components should not be understood as requiring such separation or integration in the described implementations, it being understood that the described program components and systems can generally be integrated in a single software product, or packaged into multiple software products. The above example implementations do not define or limit the invention. Other changes, substitutions, and alterations are also possible without departing from the spirit and scope of this disclosure.

It is noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of example embodiments according to the present application. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, and it should be understood that when the terms "comprises" and/or "comprising" are used in this specification, they specify the presence of stated features, steps, operations, devices, components, and/or combinations thereof, unless the context clearly indicates otherwise.

Certain terms are used throughout the description and following claims to refer to particular products. One of ordinary skill in the art will appreciate that manufacturers may refer to a component by different names. This document does not intend to distinguish between components that differ in function but not name. In the following description and claims, the terms "including," comprising, "and" including, "are intended to be open-ended terms such that they are interpreted to mean" including, but not limited to.

While the foregoing is directed to the preferred embodiment of the present invention, it will be understood by those skilled in the art that various changes and modifications may be made without departing from the spirit and scope of the invention.