Online injection product quality control method

Abstract

The invention discloses an online injection product quality control method. The method comprises the following steps that key features of technological parameters in the product injection molding process are extracted; constructing a time sequence feature prediction model and a quality prediction model based on the extracted key features; predicting process parameters of the next period through a time sequence feature prediction module, inputting the process parameters into the quality prediction model, and further predicting the size of the product in the next period; the predicted product size is verified, if the product size is qualified, technological parameter optimization does not need to be carried out, and otherwise, global optimal technological parameters are reversely searched through an iterative algorithm and substituted into the quality prediction model to carry out product size re-prediction until the product size verification is qualified. The method has the beneficial effects that by predicting the size of the injection molding product and reversely deducing the corresponding process characteristics to serve as the process data of the current injection molding machine, real-time online correction of the process data of the injection molding machine is achieved, and therefore online control over the injection molding quality is achieved.

Description

Technical Field

[0001] The present application relates to the technical field of injection molding, and in particular to an online quality control method for injection molding products. Background Art

[0002] Injection molding is one of the most widely used methods in plastic product processing. It has the advantages of molding complex structures, producing precision products, short molding cycles, high production efficiency and easy automation. Injection molding is a process in which granular or powdered raw materials are added to the hopper of the injection molding machine. The raw materials are heated and melted into a flowing state. Under the push of the screw or piston of the injection molding machine, the raw materials enter the mold cavity through the nozzle and the mold pouring system, and are cooled, hardened and formed in the mold cavity.

[0003] Injection molding is usually set up using fixed parameters input into the machine, but as batch production proceeds, changes in raw materials, production conditions and external environment often affect process parameters and cause fluctuations in the quality of injection molded products.

[0004] Currently, in the batch manufacturing process, when product quality fluctuates, it is usually necessary to rely on manual adjustment of process parameters based on historical experience. However, this method has the problems of low adjustment efficiency and poor precision, which leads to problems such as increased scrap output, decreased production efficiency and waste of resources. Summary of the invention

[0005] One of the purposes of the present application is to provide an online control method for injection molding product quality that can solve at least one of the defects in the above-mentioned background technology.

[0006] In order to achieve at least one of the above purposes, the technical solution adopted in this application is: an online control method for injection molding product quality, comprising the following steps:

[0007] S100: Extract key features of process parameters of product injection molding process;

[0008] S200: constructing a time series feature prediction model and a quality prediction model based on the extracted key features;

[0009] S300: predicting the process parameters of the next cycle through the time series feature prediction module and inputting them into the quality prediction model, thereby predicting the size of the product in the next cycle;

[0010] S400: Verify the predicted product size. If the product size is qualified, there is no need to optimize the process parameters. Otherwise, reverse search the global optimal process parameters through an iterative algorithm and substitute them into the quality prediction model to re-predict the product size until the product size verification is qualified.

[0011] S500: Adjust the current process parameters of the injection molding machine based on the corresponding process parameters when the product size verification is qualified.

[0012] Preferably, step S100 includes the following specific processes:

[0013] S110: Obtain the contribution of each feature in the process parameters to the product quality, and measure the correlation between the process parameters;

[0014] S120: Feature selection based on correlation, calculating the correlation between features and product quality and between features, and then constructing the optimal feature subset as the key feature to be extracted.

[0015] Preferably, when performing step S100, sequential forward selection is used as a feature search strategy, correlation-based feature selection is used as a performance evaluation criterion, and the Pearson correlation coefficient is used as a correlation criterion to extract key features.

[0016] Preferably, the timing feature prediction model includes an input layer, a feature processing layer, a training and testing layer, a convolutional feature extraction layer, a mixed domain attention layer and an output layer; the input layer is used for the original data input of the key process; the feature processing layer uses an algorithm to decompose the timing information of the key process features, obtains the process features of different time dimensions, and uses the original key features and the decomposed features as input; the training and testing layer extracts the timing features through the offline training module and the online reasoning module; the convolutional feature extraction layer deeply mines the local feature capabilities of nonlinear process parameters through the convolution module; the mixed domain attention layer is suitable for adaptively selecting the timing relationship of process features; the output layer is used to output the prediction results of injection molding process parameters.

[0017] Preferably, the feature processing layer includes the following specific working process: constructing an envelope curve through the local extreme points and mean line of the EMD algorithm, and then extracting the nonlinear and non-stationary features in the original data, obtaining smooth eigenfunctions with different scales, stationarities and periodic fluctuations and residuals representing the overall trend of the original data.

[0018] Preferably, the mixed domain attention layer includes a channel attention module and a spatial attention module; the channel attention module uses global average pooling and global maximum pooling to aggregate and transform feature channels to capture important temporal features; the spatial attention module is used to capture the relationship between feature temporal sequences to measure and adaptively weight the spatial relationships between different channels.

[0019] Preferably, in step S400, a reverse search of the global optimal process parameters is performed using an improved dung beetle optimization algorithm.

[0020] Preferably, the improvement of the dung beetle optimization algorithm includes the following processes: initializing the population based on Circle chaos mapping, dynamically adapting the weights based on a nonlinear convergence factor function, and randomly perturbing the update of the target position by fusing Cauchy mutation and reverse learning strategies.

[0021] Preferably, the improved formula of Circle chaos mapping is as follows:

[0022]

[0023] Among them, x n and x n+1 They represent the chaotic numbers corresponding to the nth and n+1th features respectively.

[0024] Preferably, when performing random perturbations to update the target position, the reverse learning strategy and the Cauchy mutation are applied alternately based on the selection probability function, specifically including the following process: if the number randomly generated by the rand() function in the reverse learning strategy is less than the value of the selection probability function, the reverse learning strategy is selected to perturb the update of the target position, otherwise the Cauchy mutation is used to perturb the update of the target position; after completing the perturbation of the target position, decide whether to update the target position by comparing the fitness values ​​of the new and old positions.

[0025] Compared with the prior art, the beneficial effects of this application are:

[0026] This application predicts the size of the injection molded product, and infers the corresponding process characteristics as the process data of the current injection molding machine based on the predicted size data that meets the quality requirements, thereby realizing real-time online correction of the injection molding machine process data, thereby achieving online control of the injection molding quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 The following is a schematic diagram of the workflow of this application.

[0028] Figure 2 Schematic diagram of the framework for online prediction and control of injection molding product quality in this application.

[0029] Figure 3 Schematic diagram of the EMD-MSR-MDAM model architecture in this application.

[0030] Figure 4 Schematic diagram of process feature decomposition using the EMD algorithm in this application.

[0031] Figure 5 Schematic diagram of the MSR training and reasoning module architecture in this application.

[0032] Figure 6 Schematic diagram of the architecture of the mixed domain attention module in this application.

[0033] Figure 7 Comparison of the population distribution of the chaotic map initialization in this application.

[0034] Figure 8 This is a sample table of process characteristic data after screening in this application.

[0035] Fig. 9 Table 1 is a data table showing the comparison results of the time series feature prediction models in this application.

[0036] Fig.10 Table 2 is the comparison result data of the time series feature prediction model in this application.

[0037] Fig.11 Table 3 is the data table showing the comparison results of the time series feature prediction models in this application.

[0038] Fig.12 This is a table of ablation comparison results of the time series feature prediction model in this application.

[0039] Fig.13 This is a table of hyperparameter settings for the quality prediction model in this application.

[0040] Fig.14 This is a comparison table of the quality prediction model in this application and other models.

[0041] Fig.15 This is a comparison table of product size results before and after control in this application. DETAILED DESCRIPTION

[0042] Below, the present application is further described in conjunction with specific implementation methods. It should be noted that in the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms should not be understood as necessarily being directed to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification.

[0043] In the description of the present application, it should be noted that directional words, such as the terms "center", "lateral", "longitudinal", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise", etc., indicating directions and positional relationships are based on the directions or positional relationships shown in the accompanying drawings, which are only for the convenience of narrating the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and cannot be understood as limiting the specific scope of protection of the present application.

[0044] It should be noted that the terms "first", "second", etc. in the description and claims of the present application are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence.

[0045] In this application, unless otherwise clearly specified and limited, the terms "installed", "connected", "connected", "fixed" and the like should be understood in a broad sense, for example, it can be connected, detachably connected, or integrated; it can be mechanically connected or electrically connected; it can be directly connected or indirectly connected through an intermediate medium, it can be the internal connection of two elements or the interaction relationship between two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0046] In the present application, unless otherwise clearly specified and limited, a first feature being "above" or "below" a second feature may include that the first and second features are in direct contact, or may include that the first and second features are not in direct contact but are in contact through another feature between them. Moreover, a first feature being "above", "above" and "above" a second feature includes that the first feature is directly above and obliquely above the second feature, or simply indicates that the first feature is higher in level than the second feature. A first feature being "below", "below" and "below" a second feature includes that the first feature is directly below and obliquely below the second feature, or simply indicates that the first feature is lower in level than the second feature.

[0047] The terms "including" and "having" and any variations thereof in the specification and claims of the present application are intended to cover non-exclusive inclusions. For example, a process, method, system, product or apparatus comprising a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to these processes, methods, products or apparatuses.

[0048] In order to facilitate the understanding of the technical solution of the present application, the problems existing in product quality control in the traditional injection molding process can be described in detail based on the specific injection molding stages.

[0049] The injection molding process of the product mainly includes the filling stage, the pressure holding stage, the cooling stage, and the demoulding stage. Each process contains a large number of process parameters and injection molding machine monitoring parameters. Among them, the process parameters are manually set machine parameters that can be adjusted by the operator, such as cooling time, injection speed, filling (injection) time, etc., which are adjustable process characteristic parameters; the injection molding machine monitoring data is the collected data deployed at the terminal sensor of the equipment, such as temperature average, cycle time, etc., which are non-adjustable process characteristic parameters.

[0050] The main problems faced when optimizing process parameters of injection molding production lines are:

[0051] (1) Parameter coupling. Since injection molding is a dynamic and continuous process, the quality fluctuations of each process will be transmitted, coupled and accumulated during the manufacturing process, thus affecting the final product quality. In addition, the coupling relationship between the processes is complex, and it is difficult to fit this complex mass-energy coupling relationship through a linear regression model. Therefore, when optimizing process parameters, it is necessary to study the coupling mechanism between the processes, construct a nonlinear function between process parameters and quality indicators, and then realize feedback adjustment of process parameters.

[0052] (2) Production timing. In the injection molding process, as time goes by, due to the interaction of multiple factors in the production process, the process parameters of each production cycle are different. At the same time, since injection molding is a continuous process, the temperature, pressure and other parameter states generated in the previous production cycle may have a certain impact on the subsequent production cycle, forming a timing correlation between the previous and next molds.

[0053] Based on the above analysis process, this application proposes an online control method for injection molding product quality, such as Figure 1 As shown, one of the preferred embodiments includes the following steps:

[0054] S100: Extract key features of process parameters of product injection molding process.

[0055] S200: Constructing a time series feature prediction model and a quality prediction model based on the extracted key features.

[0056] S300: Predict the process parameters of the next cycle through the timing feature prediction module and input them into the quality prediction model, thereby predicting the size of the product in the next cycle.

[0057] S400: Verify the predicted product size. If the product size is qualified, there is no need to optimize the process parameters. Otherwise, reverse search the global optimal process parameters through an iterative algorithm and substitute them into the quality prediction model to re-predict the product size until the product size verification is qualified.

[0058] S500: Adjust the current process parameters of the injection molding machine based on the corresponding process parameters when the product size verification is qualified.

[0059] It is understandable that in the injection molding production process, the key to ensuring product quality is to accurately control and optimize process parameters (such as temperature, pressure, injection speed, etc.) so that product attributes (such as size, weight, etc.) meet the preset quality control limits. Product quality prediction is to predict product quality based on process parameters, and process parameter optimization is to adjust and optimize process parameters based on product quality. These two types of tasks are interrelated. Therefore, in order to ensure the stability of product quality in each production cycle, the technical solution of this application predicts the key characteristic parameters of the next cycle through historical cycle data. Based on the prediction results, before quality problems occur in the product, feedback adjustment of process parameters is performed to avoid the production of defective products in a timely manner.

[0060] Specifically, the working framework of the technical solution of this application is as follows: Figure 2 As shown in the figure, first, a filtering method is used to perform feature screening on the process data of the injection molding machine to extract key features, such as cooling time, melt back pressure, filling (injection) time, melt temperature, etc. Secondly, a time series feature prediction model and a quality prediction model are constructed based on the key features. After that, the parameter features of the next cycle are predicted by the time series feature prediction model using the historical feature data. Then, the parameter features of the next cycle are used as input, and the product size is predicted using the online quality prediction model. The predicted product size is sent to the inspection stage for verification. If the product is qualified, no parameter optimization is required. If the product is unqualified, an iterative algorithm is used to optimize and adjust the process parameters until the product is qualified. Finally, the optimized process parameters are provided to the operator, and the process parameters of the injection molding machine are adjusted in real time to produce injection molded products, thereby realizing online control of product quality.

[0061] In this embodiment, the specific process of using the filtering method to extract key features in step S100 includes the following contents:

[0062] S110: Obtain the contribution of each feature in the process parameters to the product quality, and measure the correlation between the process parameters.

[0063] S120: Feature selection based on correlation, calculating the correlation between features and product quality and between features, and then constructing the optimal feature subset as the key feature to be extracted.

[0064] It is understandable that there are many specific algorithm types that can implement the above steps S110 and S120. In this embodiment, the Pearson correlation coefficient (PCC) is preferably used to give the contribution of each feature to the product quality, as well as to measure the correlation between process parameters. Specifically, the quality characteristics of injection molded products are affected by multiple features, and redundant features will complicate the structural network of the prediction model and waste computing resources. Therefore, in this embodiment, sequential forward selection (SFS) is used as the feature search strategy, correlation-based feature selection (CFS) is used as the performance evaluation standard, and the Pearson correlation coefficient (PCC) is used as the correlation standard to extract key features.

[0065] Specifically, the injection molding process feature screening is a process of selecting features in a sequential forward order while maintaining the order of the original data. It reduces the data dimension by calculating the correlation between the features and the target value, as well as the redundancy between the features. It includes two key stages:

[0066] Phase 1: The Pearson correlation coefficient (PCC) is used to effectively give the contribution of each feature to product quality and measure the correlation between process parameters. The formula for the correlation coefficient r calculated based on PCC is:

[0067]

[0068] Among them, r represents the Pearson correlation coefficient, x and y represent variables, and μ X represents the mean value of variable x, μ Y Represents the mean value of the variable y.

[0069] Phase II: Based on the first phase, correlation-based feature selection (CFS) is used to establish the optimal feature subset by calculating the correlation between features and product quality and between features. Then, based on the preset threshold, process data with high correlation between features but low correlation between features and quality indicators are eliminated. Among them, feature evaluation uses the heuristic equation shown below to evaluate each feature subset.

[0070]

[0071] In the formula, is the average correlation coefficient between the feature and the target value, is the average redundancy coefficient between features, V is the heuristic optimal value of the feature subset containing n features. The larger the V value, the higher the correlation between the feature and the category, and the higher the irrelevance between the features.

[0072] In this embodiment, for the construction of the time series feature prediction model, a multi-scale retention module (MSR) based on the RetNet model can be used, and a time series feature prediction model integrating empirical mode decomposition (EMD) and mixed domain attention module (MDAM) is proposed. By analyzing data from different time series modes, the trend of process characteristics can be predicted in advance, potential problems can be discovered early, and process parameters can be adjusted in time during the production process to ensure that product quality meets standards.

[0073] Specifically, Figure 3 As shown in the figure, the time series feature prediction model specifically includes an input layer, a feature processing layer, a training and testing layer, a convolutional feature extraction layer, a mixed domain attention layer, and an output layer. The input layer is used for the original data input of the key process; the feature processing layer uses an algorithm to decompose the time series information of the key process features, obtains the process features of different time dimensions, and uses the original key features and the decomposed features as input; the training and testing layer extracts the time series features through the offline training module and the online reasoning module; the convolutional feature extraction layer deeply mines the local feature capabilities of nonlinear process parameters through the convolution module; the mixed domain attention layer is suitable for adaptively selecting the time series relationship of process features; the output layer is used to output the prediction results of injection molding process parameters.

[0074] For ease of understanding, each layer of the time series feature prediction module will be described in detail below.

[0075] (1) For the feature processing layer, the envelope is constructed by using the local extreme points and mean line of the EMD algorithm to extract the nonlinear and non-stationary features in the original data, and obtain smooth eigenfunctions with different scales, stationarities and periodic fluctuations and residuals representing the overall trend of the original data.

[0076] It can be understood that the EMD decomposition algorithm can decompose complex nonlinear signals into a series of smooth intrinsic mode functions (IMFs) with different scales, stationarities, and periodic fluctuations and a residual (RES) representing the overall trend of the original data. In view of the characteristics of the injection molding production process data being unstable and noisy, the EMD algorithm uses the local extreme points and mean lines to construct envelopes based on the characteristics of the data itself, gradually extracts the nonlinear and non-stationary features in the data, effectively removes abnormal noise points in the original data, and decomposes the data into a series of smooth time series data, thereby more clearly presenting the potential rules and trends in the data. For ease of understanding, the following is a detailed explanation through specific examples.

[0077] Specifically, Figure 4 As shown, assuming that the input feature time series x = x 1 ,…,x |t|, the EMD module decomposes the feature time series into time series of different dimensions and merges them into the input vector Among them, d model Represents hidden dimensions. This enables the prediction model to learn the relationship between process parameters from different time dimensions, thereby enhancing the accuracy of the prediction results.

[0078] (2) For the training and testing layers; in the online control of injection molding product quality, considering that the injection molding machine process data is a time series based on the module frequency, the prediction of future module characteristics can preliminarily predict the future production status, thereby guiding the optimization of machine process parameters in subsequent production. Since the multi-scale retention module (MSR) in the RetNet model uses gated units to manage information flow to parallelize the processing of different parts of the input sequence, it can better capture the long-term dependencies in the sequence data while ensuring the real-time and accuracy of the prediction.

[0079] like Figure 5 As shown, the multi-scale retention module (MSR) is divided into a parallel and recursive dual retention mechanism, and the model can be trained offline in a parallel manner through a computing cluster, and tested online in a cyclic recursive manner, making online product quality control possible. Among them, the specific working process of the parallel retention mechanism and the recursive retention mechanism is well known to those skilled in the art, so it will not be elaborated in detail here; the recursive retention mechanism preferably adopts a recursive neural network structure.

[0080] It is understandable that in the injection molding production process, the temporal distribution between different process parameters is very complex, and using a single multi-scale retention mechanism may not be able to fully capture the relationship between different process parameters. Therefore, stacking the retention units side by side and focusing on the temporal relationship between various process parameters from multiple different dimensions can effectively improve the accuracy of model prediction.

[0081] (3) For the convolution feature extraction layer; the convolution module is used to extract features, which ensures the ability to extract local features of nonlinear process parameters while reducing the number of model parameters.

[0082] (4) For the mixed domain attention layer; e.g. Figure 6 As shown in Figure 1, the hybrid domain attention layer includes a channel attention module and a spatial attention module; the channel attention module uses global average pooling and global maximum pooling to aggregate and transform feature channels to capture important temporal features; the spatial attention module is used to capture the relationship between feature temporal sequences to measure and adaptively weight the spatial relationships between different channels.

[0083] It is understandable that due to the complexity and uncertainty of the injection molding production environment, various potential change factors will affect the time series trend of the parameters. As a result, the process characteristics of the previous and subsequent production cycles may not have a strong time series relationship. However, in the multi-scale retention (MSR) module, the D-matrix is ​​a causal mask with a fixed predefined weight factor, which is used to implement the attention mechanism of the model. This design assumes that the most recent time step is more important than the past time step, and uses a predefined exponential method to weight the past time step, resulting in a certain error fluctuation in the prediction result. Therefore, in order to fully extract the drastic fluctuations in process parameters caused by sudden changes at a certain moment in the injection molding process, a mixed domain attention layer is proposed to optimize the multi-scale retention (MSR) module to adaptively select the importance between time series.

[0084] It should be known that the mixed domain attention layer introduces improved efficient channel attention (IECA) based on the convolutional attention module (CBAM), and realizes the deep fusion of channel and spatial information by processing the channel attention and spatial attention modules in parallel. The classic RetNet model only uses nonlinear layers to extract the associated information of features, but after the introduction of the mixed domain attention module (MDAM), the model uses adaptive convolution technology after the convolution layer to perform deeper information mining on the channel dimension, and uses one-dimensional convolution technology to identify local features in parallel in the spatial dimension, so as to more accurately map the detailed features in the time series data.

[0085] Specifically, in the channel attention module, global average pooling (GAP) and global maximum pooling (GMP) are used to achieve efficient aggregation and transformation of feature channels. Global average pooling can not only expand the receptive field to obtain the overall change trend of channel feature information, but also remove invalid information through maximum pooling and extract sensitive channel information, so as to capture important temporal features as much as possible.

[0086] It should be known that when performing real-time control of the online injection molding process, it is not only necessary to ensure the prediction accuracy of the model, but also the optimization of the model's responsiveness is crucial. To this end, convolution operations are performed on the features after maximum pooling and average pooling, and the features obtained after convolution are weighted summed to obtain fusion features to further improve the processing speed and efficiency of the model.

[0087] Specifically, in terms of the spatial attention module, the attention module can help the model capture the relationship between the temporal aspects of the features more accurately to measure and adaptively weight the spatial relationships of different channels. The input vector X is aggregated by average pooling and maximum pooling along the channel direction. The results of global average pooling and global maximum pooling are merged, and then the spatial attention weight features are generated through convolution operations. Finally, the channel weight feature and the spatial weight feature matrix are element-wise multiplied, and then multiplied by the input vector X of the initial feature by position to obtain the salient feature X′.

[0088] (4) For the output layer; the output of the mixed domain attention layer is passed through the fully connected regression layer to reduce the dimension of the fused features after feature extraction to 1, completing the output of the injection molded product size prediction results.

[0089] In this embodiment, in terms of product quality prediction, there is no time series relationship between the process parameter features of the same module. For process data without time series and a small number of features, machine learning is a good direction relative to deep learning. There are many machine learning algorithms that can be used to build quality prediction models. The XGBoost algorithm has the advantages of parallel computing, optimized memory usage, and efficient processing of sparse data; therefore, when performing product quality prediction in this embodiment, the XGBoost algorithm is preferably used, so that prediction results with better accuracy and stability can be obtained than other linear models.

[0090] In this embodiment, when performing step S400, there are multiple specific algorithms that can perform reverse search for optimal process parameters, such as the Dung Beetle Optimization (DBO) algorithm. The Dung Beetle Optimization algorithm is inspired by the living habits of dung beetles, and by simulating multiple behavioral patterns of dung beetle populations, it can jointly achieve efficient exploration and development of space, and can achieve qualified process parameter configuration in the shortest time.

[0091] It should be known that although the dung beetle optimization algorithm can reversely search for the optimal process parameters in the shortest time, it also has the disadvantages of imbalanced global exploration and local development capabilities, easy to fall into local optimality, and weak global exploration capabilities. In order to accurately and quickly optimize the injection molding process parameters, the improved Circle chaotic mapping is used for initialization so that the initial initialization individuals are evenly distributed throughout the space. At the same time, an adaptive weight strategy is introduced to dynamically update the egg-laying and foraging behavior of dung beetles to improve the search efficiency of the algorithm. In addition, the Cauchy mutation and reverse learning strategies are incorporated into the dung beetle position update mechanism to effectively help the algorithm jump out of the local optimal solution and improve the individual quality after each iteration, thereby enhancing the search ability for the global optimal solution.

[0092] It is understandable that the improvement of the dung beetle optimization algorithm can be summarized as: initializing the population based on the Circle chaos map, dynamically adapting the weights based on the nonlinear convergence factor function, and randomly perturbing the target position update by integrating the Cauchy mutation and reverse learning strategy. For easy understanding, the specific improvement process will be described in detail below.

[0093] Specifically, for the initialization of the population, the random initial population easily leads to a large number of dung beetle individuals being unevenly gathered in space. Chaotic mapping is a key technology to solve this problem, which can significantly improve the diversity within the population. Among them, logistic chaotic mapping, Cubic chaotic mapping and Circle chaotic mapping are several methods that are currently widely used. Compared with other mapping methods, Circle chaotic mapping shows better results than other mappings in optimizing the uniformity of population distribution, especially in reducing the phenomenon of population edge aggregation. Therefore, in this embodiment, it is preferred to use the improved Cubic chaotic mapping to initialize the population.

[0094] What you need to know is that the traditional Circle chaos mapping formula is:

[0095]

[0096] The improved formula of Circle chaotic mapping in this embodiment is as follows:

[0097]

[0098] Among them, x n and x n+1 They represent the chaotic numbers corresponding to the nth and n+1th features respectively.

[0099] For ease of understanding, we can use specific parameters to understand. We can set the dimension n to 3000 and use the traditional Circle chaos mapping formula and the improved Circle chaos mapping formula to initialize the population. Figure 7 As shown in the figure, the frequency difference of the improved Circle chaotic mapping value is smaller, the distribution of dung beetle individuals is more balanced, and the diversity of the population is significantly enhanced, thereby improving the optimization efficiency.

[0100] Specifically, the lower and upper limits of the dung beetle's egg-laying and foraging areas change dynamically with the decrease coefficient R. The linear decreasing R value strategy cannot fully reflect the real dynamics of the optimization process. The use of a nonlinear approach can better optimize the synergistic effect of global and local exploration, thereby enhancing the accuracy of the algorithm. Therefore, this embodiment proposes an improved nonlinear convergence factor function form to dynamically adapt the weight strategy. The specific function formula is as follows:

[0101]

[0102] Where t represents the number of iterations, and Tmax represents the maximum number of iterations.

[0103] In the above function formula, the weight R is in exponential form, and changes slowly in the early stage of iteration. The reduction of the dung beetle's egg-laying and foraging area is slowed down, which improves the dung beetle's global search ability. As the number of iterations of the dung beetle population increases, the optimal dung beetle egg-laying and foraging area continues to be dynamically adjusted. When approaching the optimal solution in the later stage of iteration, the weight value decreases more and more. The dung beetle can achieve precise search around the optimal solution, reducing repeated exploration of the same space, strengthening local search capabilities, and effectively balancing the relationship between search diversity and convergence accuracy.

[0104] Specifically, in the DBO algorithm, the dung beetle position update depends on the position after each iteration. This update method selects the optimal dung beetle position by recalculating the fitness. Since the optimal position update lacks active intervention and is prone to fall into the local optimum, this embodiment integrates the Cauchy mutation and reverse learning strategies to randomly perturb the target position with a certain probability for update, which can effectively avoid falling into the local optimal solution too early.

[0105] It should be known that the reverse learning strategy is to find the corresponding reverse solution through the reverse learning mechanism based on the current solution, and then compare the two to retain the better solution. The Cauchy mutation is to introduce the Cauchy operator into the update process of the target position, and use its adjustment effect to enhance the ability of the algorithm to escape the local optimal solution. The specific working processes of the reverse learning strategy and the Cauchy mutation are well known to those skilled in the art, so they will not be elaborated in detail here.

[0106] In this embodiment, in order to enhance the optimization efficiency of the algorithm, it is necessary to alternately apply the reverse learning strategy and the Cauchy mutation perturbation under a specific probability. The dynamic random update mechanism fully integrates the two optimization strategies of expanding the search space (the reverse solution obtained by the reverse learning strategy) and optimizing the quality of the solution (mutating the high-quality solution through the Cauchy mutation operator), effectively alleviating the problem that the algorithm tends to fall into the local optimum.

[0107] The specific selection strategy includes the following process: if the number randomly generated by the rand() function in the reverse learning strategy is less than the value of the selection probability function, the reverse learning strategy is selected to perturb the update of the target position, otherwise the Cauchy mutation is used to perturb the update of the target position. Although the above two perturbation strategies can enhance the ability of the algorithm to jump out of the local space, it is impossible to determine whether the new position obtained after the perturbation mutation is better than the fitness value of the original position. Therefore, after completing the perturbation of the target position, the fitness value f(x) of the new and old positions is compared to decide whether to update the target position.

[0108] It should be known that the selection probability P s The formula for the fitness value f(x) is as follows:

[0109]

[0110]

[0111] Among them, θ represents the adjustment parameter, which can be 0.05; iter max Represents the built-in maximum iteration parameter; represents the reverse solution; Indicates the current optimal solution.

[0112] In order to facilitate the understanding of the technical solution of this application, the following comparative experiments will be conducted through specific parameters and the selection of appropriate models to verify the effectiveness of the injection molding product quality control method proposed in this application. The experiment includes three parts: process parameter prediction, product quality prediction and process parameter optimization and adjustment. First, predict future process parameters based on historical process parameters. Secondly, establish a quality prediction model based on the mapping relationship between process parameters and quality indicators. Finally, judge whether the process parameters can make the quality indicators meet the requirements based on the quality prediction model. If the predicted quality does not meet the standard, call the optimization algorithm to adjust the process parameters according to the quality standard. Given three size features (size1, size2 and size3) of each production mold product, and specify the qualified size range of the product: size1∈[299.85, 300.15], size2∈[199.925, 200.075], size3∈[199.925, 200.075].

[0113] In order to accurately predict product quality and provide real-time feedback optimization parameters, it is necessary to fully explore feature information and select feature parameters that are highly correlated with the target product size from the complex features. Delete the single value and features with a large number of missing values ​​in the molding machine status data set. Use the (SFS-CFS-PCC) method to evaluate and rank the importance of each feature, and at the same time explore the inherent correlation between features, remove one of the features with a high degree of correlation, and reduce the redundancy between data. After experimental analysis and screening, such as Figure 8 The table shows the sample table of process feature data after screening. A total of 9 features with low correlation between features and high correlation with target values ​​were selected, among which cooling time, filling (injection) time, melt end point, switching pressure, and barrel temperature are adjustable process parameters, while holding end point position, injection start point, cycle time, and temperature average are non-adjustable process parameters.

[0114] The prediction of the time series features of injection molded products is a typical regression problem. In order to evaluate the prediction performance of the proposed model, the 9 selected parameter features are used for time series prediction, and the first 30 historical production cycle data of the features are used to predict the process parameters of the next cycle. The data set is divided into two parts: training and testing. 80% of the time series data (13280 samples) are used for model training, and the remaining time series data (3320 samples) are used for testing. After completing step S300, some selected time series prediction models are compared. All models are single feature predictions. It is necessary to establish a prediction model for each feature for multi-feature parallel prediction. In order to evaluate the accuracy of the time series prediction model, (MSE) mean square error, RMSE (root mean square error), and MAE (mean absolute error) are used as evaluation indicators. The specific calculation formulas of each evaluation indicator are well known to those skilled in the art, so they are not elaborated in detail here.

[0115] In order to judge the performance of the time series model prediction model (EMD-MSR-MDAM), it is compared with time series models such as Attention-BLSTM, LST M-TCN, CNN-LSTM, and BiGRU to verify the effectiveness of the model. The predicted values ​​of each feature of different prediction models are as follows Figures 9 to 11 The data table shown in the figure shows that the timing prediction model of this application achieves the highest prediction accuracy in terms of switching pressure, temperature mean, injection start point, and cycle time parameters. Although the prediction accuracy of the barrel temperature, melt end point, cooling time, holding end point position, and filling time parameters is not optimal, it can be seen that the prediction results of the model in this paper still achieve good results.

[0116] At the same time, in order to verify the rationality and effectiveness of the EMD-MSR-MDAM network structure, this paper takes barrel temperature and switching pressure as prediction targets, establishes ablation experiments of MSR network, EMD-MSR network, MSR-MDAM network and EMD-MSR-MDAM network as comparison models, and trains the above models under the same experimental conditions. The experimental results are as follows Fig.12The data table shown shows a comprehensive comparison of the three main prediction indicators. The prediction effect of EMD-MSR is better than that of MSR, indicating that the EMD structure can decompose a single feature into features of multiple time scales, and can more effectively learn the correlation between the time series before and after the injection molding manufacturing production line. The prediction effect of MSR-MDAM is better than that of MSR. The MSR-MDAM network extracts the intrinsic connection between the time series relationship and the characteristics of different channels by introducing a mixed attention module, extracts the internal characteristics of the process data more comprehensively, and mines the potential laws between the time series data. Therefore, the prediction model structure of this application considers the effective combination of the three to propose the EMD-MSR-MDAM model. Experiments show that the prediction effect of the EMD-MSR-MDAM network is better than that of the MSR-MDAM and EMD-MSR networks.

[0117] From the above content, it can be seen that in terms of barrel temperature parameter prediction, the mean square error MSE, root mean square error RMSE and mean absolute error MAE of the EMD-MSR-MDAM network are reduced by 29.34%, 5.39% and 9.78% respectively compared with the MAE of the MSR network, EMD-MSR network and MSR-MDAM network. The MAE of the EMD-MSR-MDAM network is reduced by 45.18%, 12.08% and 49.03% respectively compared with the LSTM-TCN network, BiGRU network and CNN-LSTM network. In terms of switching pressure parameter prediction, the mean square error MSE, root mean square error RMSE and mean absolute error MAE of the EMD-MSR-MDAM network are reduced by 35.9%, 14.2% and 23.03% respectively compared with the MSR network, EMD-MSR network and MSR-MDAM network. The MAE of the EMD-MSR-MDAM network is reduced by 10.68%, 3.85%, 4.27%, and 4.47% respectively compared with the Attention-BLSTM network, LSTM-TCN network, BiGRU network, and CNN-LSTM network. The conclusion shows that the model of this application has smaller error and higher stability than other time series models.

[0118] In this embodiment, after completing step S300, in order to verify the effectiveness of the quality prediction model, the injection molding data set is used for model verification, which is divided into two parts: training and testing. The division of the training set and the test set and the model evaluation criteria used are the same as those of the above-mentioned time series model. XGBoost is used to compare quality predictions with artificial neural networks (ANN), LightGBM, random forest algorithms, and support vector regression (SVR). The hyperparameters of each model are optimized and set using the grid search method. The hyperparameters of the XGBoost model are as follows: Fig.13In the data table shown, for the convenience of description, the XGBoost model can be simply described as the XGB model.

[0119] pass Fig.14 The comparison results table shows that in the model comparison, the XGB model shows the best prediction effect in all three dimensions, and the root mean square errors of the XGB model in the three dimensions are 0.000375mm, 0.000281mm, and 0.000318mm respectively. The experimental results show that the XGB model has higher accuracy and stability in dealing with injection molded part size prediction problems.

[0120] In this embodiment, after completing step S400, in order to verify the accuracy of process parameter tuning, 9 injection molding state characteristics are screened out by the (SFS-CFS-PCC) method, including 5 adjustable process parameter characteristics and 4 non-adjustable process parameter characteristics. For the adjustable process characteristics, the maximum and minimum values ​​of the characteristics in the data set are taken as the adjustment optimization range of the characteristics, and the non-adjustable characteristics are not adjusted by the optimization algorithm. The total number of existing tested products is 3320, of which 721 are unqualified products. The IDBO algorithm is used to accurately adjust these process variables, and the following is obtained: Fig.15 The comparison results of the product size control before and after are shown in the table. Under the adjustment of the IDBO algorithm, most of the products that originally exceeded the specification threshold were corrected and the product size returned to the qualified range. The unqualified rates of injection molded products in the three dimensions were reduced by 59.7%, 84.5%, and 100% respectively, and the total unqualified rate of injection molded products was reduced by 85.2%.

[0121] The above describes the basic principles, main features and advantages of the present application. Those skilled in the art should understand that the present application is not limited by the above embodiments, and the above embodiments and the specification only describe the principles of the present application. The present application may have various changes and improvements without departing from the spirit and scope of the present application, and these changes and improvements fall within the scope of the present application for which protection is sought. The scope of protection claimed by the present application is defined by the attached claims and their equivalents.

Claims

1. A method for online quality control of injection molding products, characterized in that: The steps include: S100: Extract key features of process parameters of product injection molding process; S200: constructing a time series feature prediction model and a quality prediction model based on the extracted key features; S300: predicting the process parameters of the next cycle through the time series feature prediction module and inputting them into the quality prediction model, thereby predicting the size of the product in the next cycle; S400: Verify the predicted product size. If the product size is qualified, there is no need to optimize the process parameters. Otherwise, reverse search the global optimal process parameters through an iterative algorithm and substitute them into the quality prediction model to re-predict the product size until the product size verification is qualified. S500: Adjust the current process parameters of the injection molding machine based on the corresponding process parameters when the product size verification is qualified.

2. The method for online quality control of injection molding products according to claim 1, characterized in that: Step S100 includes the following specific processes: S110: Obtain the contribution of each feature in the process parameters to the product quality, and measure the correlation between the process parameters; S120: Feature selection based on correlation, calculating the correlation between features and product quality and between features, and then constructing the optimal feature subset as the key feature to be extracted.

3. The method for online quality control of injection molding products according to claim 2, characterized in that: When performing step S100, sequential forward selection is used as a feature search strategy, correlation-based feature selection is used as a performance evaluation criterion, and the Pearson correlation coefficient is used as a correlation criterion, thereby extracting key features.

4. The method for online quality control of injection molding products according to claim 1, characterized in that: The time series feature prediction model includes: Input layer: The input layer is used for the original data input of key processes; Feature processing layer: The feature processing layer uses algorithms to decompose the time series information of key process features, obtains process features of different time dimensions, and uses the original key features and the decomposed features as input; Training and testing layer: The training and testing layer extracts time series features through offline training modules and online reasoning modules. Convolutional feature extraction layer: The convolutional feature extraction layer deeply mines the local feature capabilities of nonlinear process parameters through the convolution module; A mixed domain attention layer; the mixed domain attention layer is suitable for adaptively selecting the temporal relationship of process features; and Output layer: The output layer is used to output the injection molding process parameter prediction results.

5. The method for online quality control of injection molding products according to claim 4, characterized in that: The feature processing layer includes the following specific working process: constructing the envelope curve through the local extreme points and mean line of the EMD algorithm, and then extracting the nonlinear and non-stationary features in the original data, obtaining smooth eigenfunctions with different scales, stationarities and periodic fluctuations and residuals representing the overall trend of the original data.

6. The method for online quality control of injection molding products according to claim 4, characterized in that: The mixed-domain attention layer consists of: Channel attention module; the channel attention module uses global average pooling and global maximum pooling to aggregate and transform feature channels to capture important temporal features; and Spatial attention module: The spatial attention module is used to capture the relationship between feature temporal sequences to measure and adaptively weight the spatial relationships between different channels.

7. The method for online quality control of injection molding products according to any one of claims 1 to 6, characterized in that: In step S400, a reverse search of the global optimal process parameters is performed using an improved dung beetle optimization algorithm.

8. The method for online quality control of injection molding products according to claim 7, characterized in that: The improvement of the dung beetle optimization algorithm includes the following processes: initializing the population based on Circle chaotic mapping, dynamically adapting the weights based on the nonlinear convergence factor function, and randomly perturbing the target position update by integrating Cauchy mutation and reverse learning strategies.

9. The method for online quality control of injection molding products according to claim 8, characterized in that: The improved formula of Circle chaotic mapping is as follows: Among them, x n and x n+1 They represent the chaotic numbers corresponding to the nth and n+1th features respectively.

10. The method for online quality control of injection molding products according to claim 8, characterized in that: When performing random perturbations to update the target position, the reverse learning strategy and Cauchy mutation are applied alternately based on the selection probability function, which specifically includes the following process: If the number randomly generated by the rand() function in the reverse learning strategy is less than the value of the selection probability function, the reverse learning strategy is selected to perturb the update of the target position, otherwise the Cauchy mutation is used to perturb the update of the target position; After completing the disturbance of the target position, the fitness values ​​of the new and old positions are compared to decide whether to update the target position.

Images