Adaptive plastic part wall thickness optimization control method, device and storage medium

Abstract

The present invention discloses an adaptive plastic part wall thickness optimization control method, device and storage medium, which relate to the field of plastic production detection technology. According to a pre-trained feature association network model, the low-dimensional magnetic field feature vector is cross-mapped with a preset wall thickness state library to obtain wall thickness state characteristics and determine the wall thickness fluctuation type; according to the wall thickness fluctuation type, the radial micro-displacement adjustment of the mold core is triggered, and the magnetic field signal and the mold core position data are alternately collected in real time, and the real-time wall thickness response characteristics are obtained through interactive fusion processing; based on the second autoencoder model, the difference between the real-time wall thickness response characteristics and the previous wall thickness state characteristics is dynamically reconstructed to generate wall thickness trend prediction characteristics; the trend prediction characteristics and the real-time wall thickness response characteristics are input into a reinforcement learning decision network to generate wall thickness optimization control instructions, and the temperature and mold core displacement parameters are controlled in real-time linkage.

Description

Technical Field

[0001] The present invention relates to the technical field of plastic production detection, and in particular to a method, device and storage medium for adaptively optimizing and controlling the wall thickness of plastic parts. Background Art

[0002] With the rapid development of modern industry, plastic products, especially plastic pipes, have been widely used in industrial, civil, and engineering fields due to their advantages such as light weight, corrosion resistance, and relatively low manufacturing costs. To ensure the product quality of plastic pipes, controlling the uniformity of wall thickness is particularly important.

[0003] Currently, online dynamic adjustment of mold or process parameters is primarily based on feedback from real-time sensor detection data, combined with automatic control equipment. For example, patent CN116412746B proposes "an online detection device, production line, and detection method for plastic pipe wall thickness," which detects local changes in plastic pipe wall thickness through changes in magnetic field signals, achieving real-time online detection of plastic pipe wall thickness. However, the following issues remain:

[0004] First, data collected by a single sensor typically only undergoes simple linear or low-dimensional feature analysis, making it difficult to deeply explore the information hidden within the signal and timely capture fluctuations in plastic wall thickness within a small range. Second, traditional wall thickness control systems have a relatively slow real-time response speed, making it difficult to quickly and accurately adjust process parameters. This often results in uneven localized wall thickness in the product, and may even lead to defective or scrapped products, resulting in reduced production efficiency and increased costs.

[0005] At the same time, while some solutions attempt to add multiple sensor types to obtain more comprehensive monitoring information to address these issues, they face the challenges of high complexity and significantly increased costs associated with fusing data from different sensors. Furthermore, even when multi-sensor fusion is achieved, it often involves simple overlay rather than true, in-depth data fusion analysis, failing to fundamentally improve data utilization efficiency.

[0006] In view of this, the present invention provides an adaptive plastic part wall thickness optimization control method, device and storage medium, which deeply mine the characteristic information in the sensor signal without significantly increasing the number of sensors and system cost. Summary of the Invention

[0007] The purpose of the present invention is to provide an adaptive plastic part wall thickness optimization control method, device and storage medium, which solves the problem that small fluctuations in wall thickness are difficult to accurately identify and real-time dynamic closed-loop control are difficult in existing plastic part wall thickness detection and control methods.

[0008] In order to achieve the above object, the present invention provides the following technical solutions:

[0009] In a first aspect, the present invention provides a method for adaptively optimizing and controlling the wall thickness of a plastic part, comprising the following steps:

[0010] The magnetic field signals at different positions of the target plastic part are collected and processed using wavelet transform to obtain the corresponding wavelet coefficients. The low-dimensional magnetic field feature vector is obtained by dimensionality reduction using the pre-trained first autoencoder model.

[0011] According to the pre-trained feature association network model, the low-dimensional magnetic field feature vector is cross-mapped with a preset wall thickness state library to obtain the wall thickness state feature and determine the wall thickness fluctuation type;

[0012] The radial micro-displacement of the mold core is adjusted according to the wall thickness fluctuation type, and the magnetic field signal and mold core position data are collected alternately in real time. The real-time wall thickness response characteristics are obtained through interactive fusion processing.

[0013] The second autoencoder model dynamically reconstructs the difference between the real-time wall thickness response characteristics and the previous wall thickness state characteristics to generate wall thickness trend prediction features;

[0014] The trend prediction features and the real-time wall thickness response features are input into a reinforcement learning decision network to generate wall thickness optimization control instructions, and the temperature and mold core displacement parameters are controlled in real-time linkage.

[0015] As a preferred technical solution of the first aspect of the present invention, the logic for obtaining the low-dimensional magnetic field characteristic vector is:

[0016] Acquire the magnetic field signal collected by the magnetic field sensor and process the magnetic field signal according to the wavelet transform method to obtain the corresponding wavelet coefficient; combine the wavelet coefficients of multiple magnetic field sensors to obtain a high-dimensional wavelet coefficient feature vector;

[0017] The combined processing includes:

[0018] Select the wavelet coefficients collected by multiple magnetic field sensors at the same time to construct the original wavelet coefficient matrix;

[0019] The original wavelet coefficient matrix is ​​subjected to standardization preprocessing, specifically: the wavelet coefficients corresponding to each magnetic field sensor in the matrix are respectively subjected to mean zeroing and variance standardization;

[0020] Perform feature fusion processing on the standardized wavelet coefficient matrix, and use the pre-built autoencoder network to nonlinearly map the standardized wavelet coefficient matrix into a low-dimensional feature representation;

[0021] For the low-dimensional feature vector after fusion processing, the low-dimensional magnetic field feature vector with the highest sensitivity to the wall thickness change of the plastic part is screened out through the feature selection method to obtain the optimal low-dimensional expression of the high-dimensional wavelet coefficient feature vector.

[0022] As a preferred technical solution of the first aspect of the present invention, the first autoencoder model includes an encoder and a decoder, specifically including:

[0023] The encoder is used to map the input high-dimensional wavelet coefficient feature vector to a low-dimensional magnetic field feature vector;

[0024] The decoder is used to reconstruct the low-dimensional magnetic field feature vector back into a high-dimensional wavelet coefficient feature vector;

[0025] During the training process, the high-dimensional wavelet coefficient eigenvector is used as the model input and training target, and the network parameters are optimized through the error back propagation algorithm to minimize the reconstruction error between the input and the reconstructed eigenvector.

[0026] As a preferred technical solution of the first aspect of the present invention, the specific network structure of the first autoencoder model includes:

[0027] Input layer, used to receive input high-dimensional wavelet coefficient feature vector;

[0028] The encoder sequentially sets a multi-layer fully connected network structure with the number of nodes in each layer gradually decreasing, and obtains the low-dimensional magnetic field feature vector in the middle hidden layer;

[0029] The decoder is symmetrical with the encoder structure. It sets up a multi-layer fully connected network structure in sequence, with the number of nodes in each layer gradually increasing, and outputs the reconstructed low-dimensional magnetic field feature vector.

[0030] As a preferred technical solution of the first aspect of the present invention, the application logic of the feature association network model is:

[0031] The feature association network model includes an input layer, a hidden layer and an output layer, wherein the input layer receives a low-dimensional magnetic field feature vector, and the output layer outputs an initial feature of the wall thickness state corresponding to the wall thickness state library;

[0032] By collecting magnetic field signal data of multiple historical plastic part samples, wavelet transform and autoencoder are used to obtain the corresponding low-dimensional magnetic field feature vectors as well as the actual wall thickness state data and wall thickness fluctuation type classification data of multiple historical plastic part samples;

[0033] The low-dimensional magnetic field feature vectors of the historical plastic part samples and the actual wall thickness state data are combined into a training data set to train the feature association network model;

[0034] The feature association network model is trained. During training, the low-dimensional magnetic field feature vectors of historical plastic part samples are used as network inputs, and the actual wall thickness fluctuation type is used as the training label. The error back propagation algorithm is used to train the feature association network model. The optimization goal is to minimize the loss between the predicted output and the actual wall thickness fluctuation type.

[0035] During the training process, the model weights are dynamically adjusted until the classification accuracy of the initial features of the wall thickness state output by the model reaches above the preset accuracy threshold; and the trained feature association network model is output.

[0036] As a preferred technical solution of the first aspect of the present invention, the acquisition logic of the real-time wall thickness response feature is:

[0037] According to the preset wall thickness fluctuation type mark, the mold core micro-displacement device installed inside the plastic molding mold is triggered to perform radial micro-displacement adjustment;

[0038] During the adjustment process of the mold core micro-displacement device, the magnetic field signal data detected by the magnetic field sensor arranged along the circumference of the mold outlet and the actual radial displacement data of the mold core detected by the position sensor are collected alternately in real time;

[0039] Based on the core position data, a data analysis window centered on the current core position is constructed to intercept the corresponding short-term magnetic field signal sequence;

[0040] The short-term magnetic field signal sequence and the current core position data are simultaneously input into the cross-attention network model for interactive fusion processing, wherein the core position data is used as the query vector, and the magnetic field signal sequence is used as the key vector and value vector to calculate the dynamic interaction weight;

[0041] The interactive fusion feature is output according to the dynamic interaction weight, and the interactive fusion feature is subjected to feature normalization processing to obtain the real-time wall thickness response feature of the real-time wall thickness state of the current plastic part.

[0042] As a preferred technical solution of the first aspect of the present invention, the acquisition logic of the wall thickness trend prediction feature is:

[0043] Constructing and training a second autoencoder model; the second autoencoder model includes an encoder and a decoder, wherein the encoder is used to perform nonlinear dimensionality reduction processing on the difference features between the real-time wall thickness response features and the initial features of the previous wall thickness state to obtain a low-dimensional difference feature vector; and the decoder is used to nonlinearly reconstruct the low-dimensional difference feature vector into a reconstructed difference feature vector with the same dimension as the input difference feature vector;

[0044] When training the second autoencoder model, historical wall thickness difference feature samples calculated by the difference between the real-time wall thickness response features at consecutive moments in historical plastic part production and the corresponding initial wall thickness state features are used as input and target. An error backpropagation algorithm is used, with the optimization goal of minimizing the mean square error between the input and the reconstructed output.

[0045] During the real-time production process, a current difference feature vector between the current real-time wall thickness response feature and the previous wall thickness state initial feature is obtained through differential calculation, and the difference feature vector is input into the pre-trained second autoencoder model. After the encoder dimensionality reduction processing, a low-dimensional difference feature vector is obtained as the wall thickness trend prediction feature;

[0046] The wall thickness trend prediction feature is matched with a preset historical wall thickness trend feature library, and the wall thickness trend prediction feature is output according to the matching result.

[0047] As a preferred technical solution of the first aspect of the present invention, the application logic of the wall thickness optimization control instruction is:

[0048] Construct and pre-train a reinforcement learning decision network, comprising an input layer, a policy network, and a value network. The input layer simultaneously receives both wall thickness trend prediction features and real-time wall thickness response features. The policy network outputs radial displacement parameters of the mold core micro-displacement device and temperature adjustment parameters of the temperature control device. The value network evaluates the long-term stability of the policy network output action.

[0049] The reinforcement learning decision network training method comprises: collecting wall thickness trend prediction features and real-time wall thickness response features during the historical plastic part production process as input states, recording the corresponding mold core radial displacement adjustment parameters and temperature adjustment parameters, and the corresponding actual wall thickness change results as feedback rewards; based on the state-action-feedback reward triplet, the network is trained with wall thickness stability as the goal;

[0050] During the actual plastic part production process, the real-time wall thickness response characteristics and the corresponding wall thickness trend prediction characteristics are simultaneously input into the reinforcement learning decision network to generate wall thickness optimization control instructions in real time. Based on the wall thickness optimization control instructions, the radial displacement parameters of the mold core micro-displacement device and the temperature parameters during the plastic part production process are adjusted in real time to achieve real-time closed-loop linkage control.

[0051] In a second aspect, the present invention provides an adaptive plastic part wall thickness optimization control device, based on the implementation of the first aspect, comprising a magnetic field signal feature extraction module, a wall thickness state recognition module, a real-time wall thickness response feature generation module, a wall thickness trend prediction module, and a reinforcement learning decision control module, wherein data is transmitted between the modules via wired and / or wireless communication.

[0052] The magnetic field signal feature extraction module collects magnetic field signals at different locations of the target plastic part, processes the magnetic field signals at different locations using wavelet transform to obtain corresponding wavelet coefficients, and then reduces the dimension through the pre-trained first autoencoder model to obtain a low-dimensional magnetic field feature vector;

[0053] A wall thickness state recognition module, which cross-maps the low-dimensional magnetic field feature vector with a preset wall thickness state library based on a pre-trained feature association network model to obtain wall thickness state features and determine the type of wall thickness fluctuation;

[0054] The real-time wall thickness response feature generation module triggers radial micro-displacement adjustment of the mold core according to the wall thickness fluctuation type, while alternately collecting magnetic field signals and mold core position data in real time, and obtains the real-time wall thickness response feature through interactive fusion processing;

[0055] The wall thickness trend prediction module dynamically reconstructs the difference between the real-time wall thickness response characteristics and the previous wall thickness state characteristics based on the second autoencoder model to generate wall thickness trend prediction features;

[0056] The reinforcement learning decision control module inputs the trend prediction characteristics and the real-time wall thickness response characteristics into the reinforcement learning decision network, generates wall thickness optimization control instructions, and performs real-time linkage control on the temperature and mold core displacement parameters.

[0057] In a third aspect, the present invention provides a computer-readable storage medium storing a computer program, wherein the computer program implements the steps of the first aspect when executed by a processor.

[0058] In the above technical solution, the technical effects and advantages provided by the present invention are:

[0059] The present invention combines the magnetic field signal feature extraction module with wavelet transform and the first autoencoder model to achieve efficient extraction and accurate identification of key wall thickness change features in the magnetic field signal, effectively solving the problem that traditional single methods are difficult to accurately detect tiny wall thickness fluctuations.

[0060] The present invention uses a feature association network model through a wall thickness state recognition module to cross-map the magnetic field feature vector with the wall thickness state library, accurately determine the wall thickness fluctuation type, and significantly improve the accuracy of wall thickness state recognition.

[0061] The present invention realizes real-time dynamic monitoring and instant feedback of the wall thickness status of plastic parts through the interactive fusion technology of the magnetic field signal of the real-time wall thickness response feature generation module and the mold core position data, effectively overcoming the defects of slow response speed and insufficient accuracy of traditional detection methods.

[0062] The present invention improves the real-time performance and accuracy of wall thickness trend prediction through dynamic feature reconstruction of the wall thickness trend prediction module based on the second autoencoder model, and solves the problem of poor dynamic correlation of traditional wall thickness trend prediction methods.

[0063] The present invention uses a reinforcement learning decision control module to achieve real-time closed-loop linkage control of temperature control parameters and mold core radial displacement parameters, significantly improving the accuracy, stability and coordination of plastic part wall thickness optimization control, and effectively solving the problem that traditional multi-process parameters are difficult to coordinate and optimize. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments described in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0065] Figure 1 This is a flow chart of the adaptive plastic part wall thickness optimization control method of the present invention.

[0066] Figure 2 This is a schematic plan view of the plastic part wall thickness detection method according to the present invention.

[0067] Figure 3 This is a framework diagram of the adaptive plastic part wall thickness optimization control system of the present invention. DETAILED DESCRIPTION

[0068] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in a variety of forms and should not be construed as limited to the examples set forth herein; rather, these example embodiments are provided so that the description of this disclosure will be more comprehensive and complete, and will fully convey the concepts of the example embodiments to those skilled in the art. The accompanying drawings are merely schematic illustrations of the present disclosure and are not necessarily drawn to scale. Identical reference numerals in the figures indicate identical or similar parts, and thus repeated descriptions thereof will be omitted.

[0069] In addition, the described features, structures or characteristics can be combined in any suitable manner in one or more example embodiments. In the following description, many specific details are provided to provide a full understanding of the example embodiments of the present disclosure. However, those skilled in the art will appreciate that the technical solutions of the present disclosure can be practiced while omitting one or more of the specific details, or methods, components, steps, etc. In other cases, well-known structures, methods, implementations or operations are not shown or described in detail to avoid obscuring various aspects of the present disclosure.

[0070] Example 1

[0071] like Figure 1 As shown, the present invention provides an adaptive plastic part wall thickness optimization control method, comprising the following steps:

[0072] S101: Collect magnetic field signals at different positions of the plastic part, process each magnetic field signal according to the wavelet transform method to obtain corresponding wavelet coefficients; and perform dimensionality reduction processing on the wavelet coefficients according to the pre-trained first autoencoder model to obtain a low-dimensional magnetic field feature vector;

[0073] It should be noted that the use of magnetic field sensors combined with wavelet transform methods to accurately extract signal features, and the use of autoencoder models to perform fine dimensionality reduction to extract effective features, solves the problem that traditional single methods cannot achieve accurate detection of tiny wall thickness fluctuations.

[0074] Specifically, the logic for obtaining the low-dimensional magnetic field feature vector is:

[0075] Acquire the magnetic field signal collected by the magnetic field sensor and process the magnetic field signal according to the wavelet transform method to obtain the corresponding wavelet coefficient; combine the wavelet coefficients of multiple magnetic field sensors to obtain a high-dimensional wavelet coefficient feature vector;

[0076] The combined processing includes:

[0077] Select the wavelet coefficients collected by multiple magnetic field sensors at the same time to construct the original wavelet coefficient matrix;

[0078] The original wavelet coefficient matrix is ​​subjected to standardization preprocessing, specifically: the wavelet coefficients corresponding to each magnetic field sensor in the matrix are respectively subjected to mean zeroing and variance standardization;

[0079] Perform feature fusion processing on the standardized wavelet coefficient matrix, and use the pre-built autoencoder network to nonlinearly map the standardized wavelet coefficient matrix into a low-dimensional feature representation;

[0080] For the low-dimensional feature vector after fusion processing, the low-dimensional magnetic field feature vector with the highest sensitivity to the wall thickness change of the plastic part is screened out through the feature selection method to obtain the optimal low-dimensional expression of the high-dimensional wavelet coefficient feature vector.

[0081] Constructing a first autoencoder model; wherein the first autoencoder model includes an encoder and a decoder; the encoder is used to map the input high-dimensional wavelet coefficient feature vector to obtain a low-dimensional magnetic field feature vector with a dimension lower than the input feature vector; the decoder is used to map the low-dimensional magnetic field feature vector to reconstruct a reconstructed feature vector with the same dimension as the input feature vector;

[0082] Training the first autoencoder model; during training, using a high-dimensional wavelet coefficient feature vector as a model input and the high-dimensional wavelet coefficient feature vector itself as a training target; optimizing network parameters of the first autoencoder model using an error backpropagation algorithm with minimizing a reconstruction error between the input and the reconstructed feature vector as a training target;

[0083] After the training is completed, the high-dimensional wavelet coefficient feature vector corresponding to the plastic part to be inspected is input into the trained encoder, and the low-dimensional magnetic field feature vector is obtained through mapping, which is used for the subsequent identification or analysis of the wall thickness status of the plastic part.

[0084] For example, Figure 2 As shown, assuming we have 4 magnetic field sensors, then N=4: each magnetic field sensor obtains a set of magnetic field signals; each set of magnetic field signals generates a set of wavelet coefficients Xn after wavelet transformation; thus forming a high-dimensional data vector X=[X1,X2,X3,X4]:

[0085] Assuming that the length of the wavelet coefficients obtained by each magnetic field sensor after wavelet transform is: M (such as 64 dimensions, 128 dimensions, etc.), the dimension of the magnetic field data obtained at the current timestamp is: dimension = N × M dimensions. For example: if each sensor generates 64-dimensional wavelet coefficients, the dimension of the high-dimensional data vector is 256 dimensions (4 × 64).

[0086] The specific network structure of the first autoencoder model includes:

[0087] Input layer, used to receive input high-dimensional wavelet coefficient feature vector;

[0088] The encoder sequentially sets a multi-layer fully connected network structure with the number of nodes in each layer gradually decreasing, and obtains the low-dimensional magnetic field feature vector in the middle hidden layer;

[0089] The decoder is symmetrical with the encoder structure, and sets up a multi-layer fully connected network structure in sequence. The number of nodes in each layer gradually increases, and outputs the reconstructed low-dimensional magnetic field feature vector;

[0090] Taking the high-dimensional feature dimension as 512 and the low-dimensional magnetic field feature vector dimension as 8 as an example, the network structure of the first autoencoder model is set as follows:

[0091] The number of input layer nodes is 512;

[0092] The encoder sets the fully connected network layers with 128, 32 and 8 nodes respectively;

[0093] The decoder sets up fully connected network layers with 32, 128 and 512 nodes respectively;

[0094] The ReLU activation function is used between each network layer, and the output layer uses the Sigmoid or Tanh activation function.

[0095] Through the above method, the first autoencoder model achieves effective dimensionality reduction of the high-dimensional feature vector of the magnetic field signal and obtains a unified and stable low-dimensional magnetic field feature representation. It uses high-dimensional data as input and itself as the training target, and trains the autoencoder by optimizing the loss function (such as mean square error).

[0096] S102: cross-mapping the low-dimensional magnetic field feature vector with a preset wall thickness state library based on a pre-trained feature association network model to obtain initial features of the wall thickness state; and determining the wall thickness fluctuation type based on the initial features of the wall thickness state, and outputting a fluctuation type label;

[0097] It should be noted that: a feature association network model is constructed, a cross-mapping relationship between low-dimensional magnetic field features and a preset wall thickness state library is established, different types of wall thickness fluctuations are identified, and the problem of difficulty in identifying wall thickness states is effectively solved.

[0098] Specifically, the application logic of the feature association network model is:

[0099] Constructing and pre-training the feature association network model, wherein the feature association network model is used to establish a cross-mapping relationship between the low-dimensional magnetic field feature vector and the wall thickness state of the plastic part;

[0100] The feature association network model includes an input layer, a hidden layer and an output layer. The input layer receives a low-dimensional magnetic field feature vector, and the output layer outputs an initial feature of the wall thickness state corresponding to the wall thickness state library.

[0101] The wall thickness state library is established by collecting actual wall thickness measurement data of historical plastic part samples; specifically, it includes:

[0102] Collect magnetic field signal data of multiple historical plastic part samples, and use the same wavelet transform and autoencoder method as the present invention to obtain corresponding low-dimensional magnetic field feature vectors;

[0103] Use high-precision wall thickness measuring instruments to measure the actual wall thickness status data of multiple historical plastic part samples, and classify the wall thickness fluctuation types of the plastic part samples based on the measurement data;

[0104] The low-dimensional magnetic field feature vectors of the historical plastic part samples and the corresponding actual wall thickness fluctuation type classification data are combined into a training data set to train the feature association network model;

[0105] The feature association network model is trained. During training, the low-dimensional magnetic field feature vectors of historical plastic part samples are used as network inputs, and the actual wall thickness fluctuation type is used as the training label. The training process is as follows:

[0106] The training data set is randomly divided into a training set and a validation set, and the error back propagation algorithm is used to train the feature association network model;

[0107] The training goal is to accurately map the initial features of the wall thickness state output by the model to the corresponding wall thickness fluctuation type. Specifically, the cross entropy loss function is used, and the optimization goal is to minimize the loss between the model prediction output and the actual wall thickness fluctuation type.

[0108] Dynamically adjust the model weights during the training process until the classification accuracy of the model output initial features of the wall thickness state reaches above the preset accuracy threshold;

[0109] After the model training is completed, during the actual inspection process, the low-dimensional magnetic field feature vector of the plastic part to be inspected is input into the feature association network model. Through the cross-mapping processing of the model, the initial wall thickness state feature corresponding to the plastic part to be inspected is directly output;

[0110] Determining the type of wall thickness fluctuation based on the initial characteristics of the wall thickness state; specifically including:

[0111] The initial features of the wall thickness state output by the feature association network model are processed through a pre-built classification module (such as a Softmax classifier or a support vector machine). Based on the processing results of the classification module, a wall thickness fluctuation type label is output for subsequent triggering of adjustments to the mold core micro-displacement device or optimization of production process parameters.

[0112] The feature association network model effectively realizes the precise and cross-mapping between the low-dimensional magnetic field feature vector and the wall thickness state of the plastic part through the above steps, thereby accurately and efficiently obtaining the initial characteristics of the wall thickness state and realizing real-time determination of the wall thickness fluctuation type.

[0113] S103: while outputting the fluctuation type mark, triggering the mold core micro-displacement device to perform a radial micro-displacement adjustment process; and during the process of the mold core micro-displacement device performing the radial micro-displacement adjustment, alternately collecting magnetic field signals generated by the radial micro-displacement and mold core position data, and interactively fusing the magnetic field signals and the mold core position data to obtain a real-time wall thickness response feature;

[0114] It should be noted that: through real-time interactive fusion of magnetic field signals and mold core position data, an instant feedback closed loop is achieved. During the real-time adjustment of the process, the wall thickness response characteristics of the plastic part are updated in real time, effectively solving the problem of slow real-time response of traditional methods.

[0115] Specifically, during the adjustment process of the mold core micro-displacement device, the method for interactively fusing and processing the magnetic field signal and the mold core position data in real time to obtain the real-time wall thickness response characteristics specifically includes the following steps:

[0116] According to the wall thickness fluctuation type mark obtained in step S102, the mold core micro-displacement device is triggered to perform a radial micro-displacement adjustment process; the mold core micro-displacement device is specifically a piezoelectric drive device or an electromagnetic drive device, which is installed inside the plastic molding mold and is used to accurately adjust the radial position of the mold core in real time;

[0117] During the radial micro-displacement adjustment process of the mold core micro-displacement device, the following two types of data are collected in real time using an alternating acquisition method:

[0118] Real-time collection of magnetic field signal data from magnetic field sensors, where the magnetic field sensors are evenly arranged along the circumference of the mold outlet to sense changes in magnetic field intensity caused by changes in the mold core position;

[0119] At the same time, the core position data of the core micro-displacement device is collected in real time, wherein the core position data is the actual radial displacement of the core, which is obtained by a position sensor installed on the core micro-displacement device;

[0120] Specifically, the real-time alternating acquisition frequency of the magnetic field signal and the mold core position data is kept consistent, with a sampling frequency of 100Hz to 500Hz, to ensure strict temporal synchronization of the two types of data;

[0121] Based on the above real-time collected magnetic field signal data and the core position data, interactive fusion processing is performed; the specific implementation steps include:

[0122] First, using the core position data as a benchmark, a data analysis window centered on the current core position is constructed; within this window, a short-term magnetic field signal sequence corresponding to the current core position is intercepted;

[0123] Then, the short-term magnetic field signal sequence and the current core position data are input into a pre-built cross-attention network model for interactive fusion processing;

[0124] The cross attention network uses the core position data as the query vector and the magnetic field signal sequence as the key vector and value vector, and calculates the dynamic interaction weight between the two.

[0125] Specifically, the calculation formula of the cross attention network is as follows:

[0126] ;

[0127] in, represents the query vector consisting of the core position data, 、 represents the key vector and value vector composed of the magnetic field signal sequence, is the key vector dimension;

[0128] Based on the interactive fusion processing of the above-mentioned cross-attention network, an interactive fusion feature with real-time dynamics is output; the interactive fusion feature is further normalized to obtain a real-time wall thickness response feature that can accurately characterize the real-time wall thickness state of the current plastic part, which is used for subsequent wall thickness trend prediction and optimization decision analysis. The real-time interactive fusion processing between the magnetic field signal and the mold core position data is realized, which can quickly and accurately capture the real-time wall thickness response characteristics of the mold core position change to the plastic part wall thickness, and effectively solve the technical problems of slow wall thickness response and insufficient accuracy in traditional methods.

[0129] S104: While obtaining the real-time wall thickness response feature, dynamically reconstructing the difference between the current real-time wall thickness response feature and the previous wall thickness state initial feature based on the second autoencoder model to generate a wall thickness trend prediction feature; and matching the wall thickness trend prediction feature with a preset historical wall thickness trend feature library to output the wall thickness trend prediction feature;

[0130] It should be noted that the introduction of the second autoencoder model dynamically reconstructs the difference between the real-time wall thickness characteristics and the historical state characteristics, improves the accuracy of wall thickness trend prediction, and solves the problem of insufficient dynamic correlation of traditional prediction methods.

[0131] Specifically, the method of performing dynamic feature reconstruction processing based on the second autoencoder model to generate wall thickness trend prediction features and outputting the wall thickness trend prediction features accordingly specifically includes the following steps:

[0132] A second autoencoder model is constructed, and the second autoencoder model is used to realize dynamic nonlinear reconstruction of the difference characteristics between the real-time wall thickness response characteristics and the initial characteristics of the previous wall thickness state; wherein the second autoencoder model specifically includes an encoder and a decoder:

[0133] The encoder is used to perform nonlinear dimensionality reduction processing on the input difference feature vector to obtain a low-dimensional difference feature vector;

[0134] The decoder is used to nonlinearly reconstruct the low-dimensional difference feature vector and output a reconstructed difference feature vector with the same dimension as the input difference feature vector.

[0135] The training method of the second autoencoder model specifically includes:

[0136] Collect the real-time wall thickness response characteristics and the corresponding initial wall thickness characteristics at consecutive moments in the historical plastic part production process, calculate the difference between the two, and obtain the historical wall thickness difference characteristic samples;

[0137] Using the historical wall thickness difference feature samples as training input for a second autoencoder model, and using the historical wall thickness difference feature samples themselves as training targets for training;

[0138] During the training process, the error back propagation algorithm is used to optimize the network parameters of the second autoencoder model with the minimum mean square error (MSE) between the input and the reconstructed output as the optimization goal;

[0139] After training, the second autoencoder model has the ability to accurately extract the potential trend features within the wall thickness difference features;

[0140] During the real-time production process, while obtaining the current real-time wall thickness response characteristics, the current real-time wall thickness response characteristics and the initial characteristics of the previous wall thickness state are differentially processed to calculate the current difference characteristic vector;

[0141] The specific difference calculation method is:

[0142] Difference feature vector = current real-time wall thickness response feature − previous wall thickness state initial feature

[0143] The current difference feature vector obtained above is input into the pre-trained second autoencoder model, and the encoder performs nonlinear dimensionality reduction to obtain a low-dimensional difference feature vector, which is then nonlinearly reconstructed by the decoder to obtain a reconstructed difference feature vector; the low-dimensional difference feature vector output by the encoder is used as the wall thickness trend prediction feature;

[0144] Establish a historical wall thickness trend feature library, including:

[0145] During the historical production process, the second autoencoder model is used to obtain low-dimensional difference feature vectors of multiple historical plastic part samples, and the corresponding actual wall thickness change trends are recorded at the same time;

[0146] A historical wall thickness trend feature library is formed, in which each historical feature vector is marked with a corresponding wall thickness trend category or trend value;

[0147] The currently obtained wall thickness trend prediction features are matched with the historical wall thickness trend features in the historical wall thickness trend feature library; the specific matching method is:

[0148] Use the nearest neighbor algorithm or cosine similarity algorithm to calculate the distance or similarity between the current wall thickness trend prediction feature and each historical feature in the historical wall thickness trend feature library; determine the wall thickness trend category or trend change value corresponding to the historical feature with the highest matching degree;

[0149] Based on the above feature matching results, the wall thickness trend prediction feature is output. The wall thickness trend prediction feature is used for real-time optimization decision-making of subsequent plastic part wall thickness control. It realizes dynamic nonlinear feature reconstruction and trend prediction matching between the real-time wall thickness response feature and the historical wall thickness state feature. It can accurately predict the changing trend of the plastic part wall thickness and provide a reliable decision-making basis for real-time wall thickness control.

[0150] S105: The wall thickness trend prediction feature and the real-time wall thickness response feature are simultaneously input into a pre-trained reinforcement learning decision network to generate a wall thickness optimization control instruction; and according to the wall thickness optimization control instruction, the radial displacement parameters of the temperature control device and the mold core micro-displacement device in the plastic part production process are controlled in a real-time closed-loop linkage manner.

[0151] It should be noted that: through the reinforcement learning decision network, trend prediction features and real-time wall thickness response features are input simultaneously to achieve real-time linkage control of temperature parameters and core micro-displacement parameters, thereby improving the overall wall thickness adaptive control accuracy and coordination, and solving the problem of linkage control of multiple process parameters.

[0152] Specifically, the specific implementation method of simultaneously inputting the wall thickness trend prediction features and the real-time wall thickness response features into the reinforcement learning decision network to generate wall thickness optimization control instructions and perform real-time closed-loop linkage control includes:

[0153] Constructing and pre-training a reinforcement learning decision network, which is used to implement nonlinear decision mapping between plastic part wall thickness trend prediction characteristics and real-time wall thickness response characteristics to output accurate wall thickness optimization control instructions;

[0154] Specifically, the reinforcement learning decision network includes an input layer, a policy network, and a value network:

[0155] The input layer receives both wall thickness trend prediction features and real-time wall thickness response features;

[0156] The strategy network is used to output specific process parameter adjustment actions, including the radial displacement adjustment parameters of the mold core micro-displacement device and the temperature adjustment parameters during the plastic part production process;

[0157] The value network is used to evaluate the long-term stability and control effectiveness of the policy network's output actions to ensure the stability of the optimization instructions;

[0158] The training method of the reinforcement learning decision network specifically includes:

[0159] During the historical plastic part production process, a large number of wall thickness trend prediction features and real-time wall thickness response features are collected as state inputs, and the corresponding process parameter adjustment actions and the actual wall thickness changes caused by them are recorded;

[0160] Specifically, in the historical plastic part production process, the process parameter adjustment actions include the mold core radial displacement adjustment amplitude, adjustment frequency, and the temperature setting value of the temperature control device;

[0161] The state-action-feedback result triplet is used as training data for the reinforcement learning decision network; the feedback result is evaluated by the wall thickness stability or the wall thickness fluctuation range;

[0162] During training, the policy network is updated based on gradients to maximize long-term feedback rewards, while the value network is optimized with the long-term stability of the wall thickness state as the goal. Specifically, the Deep Deterministic Policy Gradient (DDPG) or Proximal Policy Optimization (PPO) algorithm is used to complete the training.

[0163] After the training is completed, during the actual plastic part production process, the current real-time wall thickness response characteristics and the corresponding wall thickness trend prediction characteristics are simultaneously input into the pre-trained reinforcement learning decision network.

[0164] The reinforcement learning decision network outputs specific wall thickness optimization control instructions, including:

[0165] Radial displacement parameter adjustment instructions for the mold core micro-displacement device; specifically including radial displacement adjustment amount, adjustment speed, and adjustment cycle;

[0166] Temperature parameter adjustment instructions for temperature control devices during plastic part production; specifically including heating or cooling temperature setpoints and temperature adjustment rates;

[0167] According to the wall thickness optimization control instruction, the temperature control device and the mold core micro-displacement device in the plastic part production process are controlled in a real-time closed-loop linkage manner; the specific implementation is as follows:

[0168] The mold core micro-displacement device adjusts the radial position of the mold core in real time according to the radial displacement parameter adjustment instruction;

[0169] The temperature control device adjusts the temperature of the molding die or hot runner area in real time according to the temperature parameter adjustment instruction;

[0170] The actual wall thickness response state of the plastic part after the above-mentioned adjustment action is collected in real time and fed back to the reinforcement learning decision network for decision adjustment in the next cycle, thereby realizing dynamic closed-loop optimization control and accurate real-time control of the wall thickness of the plastic part. It can dynamically adapt to the wall thickness fluctuations caused by material fluctuations, process fluctuations and environmental changes in the plastic part production process, and effectively improve the uniformity of the wall thickness of the plastic part and product quality.

[0171] Example 2

[0172] The parts not described in detail in this embodiment are as shown in Example 1. This embodiment is based on an exemplary description of a plastic part processing site. First, a magnetic field sensor and a mold core micro-displacement device are used on a plastic molding mold to detect the plastic part. Specifically, the following steps are included:

[0173] (1) Installation location and detection data of magnetic field sensor:

[0174] Magnetic field sensors are usually installed on the outside near the outlet of the plastic molding mold. Magnetic field sensors are evenly arranged along the mold outlet cross-section to detect the target plastic part, which can be a tubular or annular plastic part. Each magnetic field sensor is set at an appropriate distance from the surface of the plastic part (usually several millimeters to several centimeters) to form a non-contact detection structure.

[0175] Based on the magnetic field intensity variation information at different positions (along the circumference of the target plastic part), that is, the difference in magnetic field signals caused by the radial wall thickness fluctuation of the plastic part; a fixed magnetic marker assembly is installed inside the plastic part (as described in patent CN116412746B), and the magnetic field sensor can sense the change in the distance between the magnetic marker assembly and the sensor in a non-contact manner. When the wall thickness of the plastic part changes, the position of the magnetic marker assembly corresponding to the inner wall of the plastic part undergoes a slight displacement, resulting in a change in the magnetic field intensity. The change in magnetic field intensity signal can indirectly determine whether the plastic part wall thickness is uniform and the specific fluctuation state.

[0176] (2) Installation position and test data of the core micro-displacement device:

[0177] A core micro-displacement device (such as a piezoelectric actuator or electromagnetic fine-tuning device) is located within the core structure of a plastic molding mold, either at the end or in the middle adjustable section. It precisely controls the core's radial position by micro-steps. This includes: core radial position (real-time absolute or relative position); real-time displacement data (displacement magnitude, speed, and frequency) associated with core radial position adjustments; and theoretical changes in plastic part wall thickness caused by core position fine-tuning.

[0178] (3) Relationship between core displacement and wall thickness:

[0179] Slight changes in the radial position of the mold core (such as fine-tuning outward or inward) directly affect the cross-sectional size of the plastic material flow channel, thereby changing the material flow thickness and the molded wall thickness of the plastic part. Real-time collection of mold core position data can confirm the current trend of plastic part wall thickness changes and the accuracy of adjustment effects.

[0180] In summary, externally detected magnetic field sensor data reflects changes in part wall thickness in real time, while internally detected core position data reflects the precise implementation of wall thickness adjustments. Interactive fusion technology allows for real-time comparison and integration of these data, enabling precise closed-loop control. Real-time magnetic field sensor data serves as feedback to determine the current actual wall thickness, while core position data serves as input and feedback for the actuator's real-time adjustments, verifying that adjustments align with the actual part wall thickness. This approach not only improves the accuracy and real-time nature of wall thickness detection and control, but also avoids the cost of additional sensors required in traditional methods.

[0181] During the initial plastic part production phase, magnetic field signals and initial core position data are collected for a batch of parts to establish an initial calibration database. Using a high-precision thickness gauge or coordinate measuring machine, the actual wall thickness data for each batch of plastic parts is accurately measured to form an initial wall thickness status database. Using the collected magnetic field data and core position data as input and the actual wall thickness data as supervision, a cross-attention network model is trained to learn interactive fusion rules. After model convergence, the pre-trained model is solidified and deployed on the on-site industrial computer for backup.

[0182] During the actual production process of plastic parts, the real-time position signals from the mold core micro-displacement device and the magnetic field sensor are alternately transmitted to the aforementioned edge computing module in real time. At the beginning of each production cycle, magnetic field signals at different locations on the plastic part and the real-time position data of the mold core are collected in real time. Alternating intervals (for example, 100Hz) are used to ensure accurate time alignment between the magnetic field data and the mold core position data.

[0183] The edge computing unit executes the deployed cross-attention network algorithm in real time, quickly completing the interactive fusion of magnetic field signals and mold core position data to obtain real-time wall thickness response characteristics. Based on the real-time wall thickness response characteristics, the on-site edge computing module determines the gap between the actual state of the current plastic part wall thickness and the preset standard; and outputs the mold core micro-displacement device adjustment instructions and temperature control equipment adjustment instructions in real time through the reinforcement learning decision module.

[0184] Example 3

[0185] like Figure 3 As shown, the parts not described in detail in this embodiment are as shown in Example 1. This embodiment provides an adaptive plastic part wall thickness optimization control device, including a magnetic field signal feature extraction module, a wall thickness state recognition module, a real-time wall thickness response feature generation module, a wall thickness trend prediction module and a reinforcement learning decision control module. Data is transmitted between each module via wired and / or wireless communication.

[0186] The magnetic field signal feature extraction module collects magnetic field signals at different locations of the target plastic part, processes the magnetic field signals at different locations using wavelet transform to obtain corresponding wavelet coefficients, and then reduces the dimension through the pre-trained first autoencoder model to obtain a low-dimensional magnetic field feature vector;

[0187] The acquisition logic of the low-dimensional magnetic field characteristic vector is:

[0188] Acquire the magnetic field signal collected by the magnetic field sensor and process the magnetic field signal according to the wavelet transform method to obtain the corresponding wavelet coefficient; combine the wavelet coefficients of multiple magnetic field sensors to obtain a high-dimensional wavelet coefficient feature vector;

[0189] The combined processing includes:

[0190] Select the wavelet coefficients collected by multiple magnetic field sensors at the same time to construct the original wavelet coefficient matrix;

[0191] The original wavelet coefficient matrix is ​​subjected to standardization preprocessing, specifically: the wavelet coefficients corresponding to each magnetic field sensor in the matrix are respectively subjected to mean zeroing and variance standardization;

[0192] Perform feature fusion processing on the standardized wavelet coefficient matrix, and use the pre-built autoencoder network to nonlinearly map the standardized wavelet coefficient matrix into a low-dimensional feature representation;

[0193] For the low-dimensional feature vector after fusion processing, the low-dimensional magnetic field feature vector with the highest sensitivity to the wall thickness change of the plastic part is screened out through the feature selection method to obtain the optimal low-dimensional expression of the high-dimensional wavelet coefficient feature vector.

[0194] The first autoencoder model includes an encoder and a decoder, specifically including:

[0195] The encoder is used to map the input high-dimensional wavelet coefficient feature vector to a low-dimensional magnetic field feature vector;

[0196] The decoder is used to reconstruct the low-dimensional magnetic field feature vector back into a high-dimensional wavelet coefficient feature vector;

[0197] During the training process, the high-dimensional wavelet coefficient eigenvector is used as the model input and training target, and the network parameters are optimized through the error back propagation algorithm to minimize the reconstruction error between the input and the reconstructed eigenvector.

[0198] The specific network structure of the first autoencoder model includes:

[0199] Input layer, used to receive input high-dimensional wavelet coefficient feature vector;

[0200] The encoder sequentially sets a multi-layer fully connected network structure with the number of nodes in each layer gradually decreasing, and obtains the low-dimensional magnetic field feature vector in the middle hidden layer;

[0201] The decoder is symmetrical with the encoder structure. It sets up a multi-layer fully connected network structure in sequence, with the number of nodes in each layer gradually increasing, and outputs the reconstructed low-dimensional magnetic field feature vector.

[0202] A wall thickness state recognition module, which cross-maps the low-dimensional magnetic field feature vector with a preset wall thickness state library based on a pre-trained feature association network model to obtain wall thickness state features and determine the type of wall thickness fluctuation;

[0203] The application logic of the feature association network model is:

[0204] The feature association network model includes an input layer, a hidden layer and an output layer, wherein the input layer receives a low-dimensional magnetic field feature vector, and the output layer outputs an initial feature of the wall thickness state corresponding to the wall thickness state library;

[0205] By collecting magnetic field signal data of multiple historical plastic part samples, wavelet transform and autoencoder are used to obtain the corresponding low-dimensional magnetic field feature vectors as well as the actual wall thickness state data and wall thickness fluctuation type classification data of multiple historical plastic part samples;

[0206] The low-dimensional magnetic field feature vectors of the historical plastic part samples and the actual wall thickness state data are combined into a training data set to train the feature association network model;

[0207] The feature association network model is trained. During training, the low-dimensional magnetic field feature vectors of historical plastic part samples are used as network inputs, and the actual wall thickness fluctuation type is used as the training label. The error back propagation algorithm is used to train the feature association network model. The optimization goal is to minimize the loss between the predicted output and the actual wall thickness fluctuation type.

[0208] During the training process, the model weights are dynamically adjusted until the classification accuracy of the initial features of the wall thickness state output by the model reaches above the preset accuracy threshold; and the trained feature association network model is output.

[0209] The real-time wall thickness response feature generation module triggers radial micro-displacement adjustment of the mold core according to the wall thickness fluctuation type, while alternately collecting magnetic field signals and mold core position data in real time, and obtains the real-time wall thickness response feature through interactive fusion processing;

[0210] The acquisition logic of the real-time wall thickness response feature is:

[0211] According to the preset wall thickness fluctuation type mark, the mold core micro-displacement device installed inside the plastic molding mold is triggered to perform radial micro-displacement adjustment;

[0212] During the micro-displacement adjustment process of the mold core, the magnetic field signal data detected by the magnetic field sensor arranged along the circumference of the mold outlet and the actual radial displacement data of the mold core detected by the position sensor are collected alternately in real time;

[0213] Based on the core position data, a data analysis window centered on the current core position is constructed to intercept the corresponding short-term magnetic field signal sequence;

[0214] The short-term magnetic field signal sequence and the current core position data are simultaneously input into the cross-attention network model for interactive fusion processing, wherein the core position data is used as the query vector, and the magnetic field signal sequence is used as the key vector and value vector to calculate the dynamic interaction weight;

[0215] The interactive fusion feature is output according to the dynamic interaction weight, and the interactive fusion feature is subjected to feature normalization processing to obtain the real-time wall thickness response feature of the real-time wall thickness state of the current plastic part.

[0216] The wall thickness trend prediction module dynamically reconstructs the difference between the real-time wall thickness response characteristics and the previous wall thickness state characteristics based on the second autoencoder model to generate wall thickness trend prediction features;

[0217] The acquisition logic of the wall thickness trend prediction feature is:

[0218] Constructing and training a second autoencoder model; the second autoencoder model includes an encoder and a decoder, wherein the encoder is used to perform nonlinear dimensionality reduction processing on the difference features between the real-time wall thickness response features and the initial features of the previous wall thickness state to obtain a low-dimensional difference feature vector; and the decoder is used to nonlinearly reconstruct the low-dimensional difference feature vector into a reconstructed difference feature vector with the same dimension as the input difference feature vector;

[0219] When training the second autoencoder model, historical wall thickness difference feature samples calculated by the difference between the real-time wall thickness response features at consecutive moments in historical plastic part production and the corresponding initial wall thickness state features are used as input and target. An error backpropagation algorithm is used, with the optimization goal of minimizing the mean square error between the input and the reconstructed output.

[0220] During the real-time production process, a current difference feature vector between the current real-time wall thickness response feature and the previous wall thickness state initial feature is obtained through differential calculation, and the difference feature vector is input into the pre-trained second autoencoder model. After the encoder dimensionality reduction processing, a low-dimensional difference feature vector is obtained as the wall thickness trend prediction feature;

[0221] The wall thickness trend prediction feature is matched with a preset historical wall thickness trend feature library, and the wall thickness trend prediction feature is output according to the matching result.

[0222] The reinforcement learning decision control module inputs the trend prediction characteristics and the real-time wall thickness response characteristics into the reinforcement learning decision network, generates wall thickness optimization control instructions, and performs real-time linkage control on the temperature and mold core displacement parameters.

[0223] The application logic of the wall thickness optimization control instruction is:

[0224] Construct and pre-train a reinforcement learning decision network, comprising an input layer, a policy network, and a value network. The input layer simultaneously receives both wall thickness trend prediction features and real-time wall thickness response features. The policy network outputs radial displacement parameters of the mold core micro-displacement device and temperature adjustment parameters of the temperature control device. The value network evaluates the long-term stability of the policy network output action.

[0225] The reinforcement learning decision network training method comprises: collecting wall thickness trend prediction features and real-time wall thickness response features during the historical plastic part production process as input states, recording the corresponding mold core radial displacement adjustment parameters and temperature adjustment parameters, and the corresponding actual wall thickness change results as feedback rewards; based on the state-action-feedback reward triplet, the network is trained with wall thickness stability as the goal;

[0226] During the actual plastic part production process, the real-time wall thickness response characteristics and the corresponding wall thickness trend prediction characteristics are simultaneously input into the reinforcement learning decision network to generate wall thickness optimization control instructions in real time. Based on the wall thickness optimization control instructions, the radial displacement parameters of the mold core micro-displacement device and the temperature parameters during the plastic part production process are adjusted in real time to achieve real-time closed-loop linkage control.

[0227] This embodiment provides an adaptive plastic part wall thickness optimization control method, which is used to execute an adaptive plastic part wall thickness optimization control device provided by the above-mentioned embodiments of the present invention. The specific methods and processes for realizing corresponding functions based on the various structures included in the adaptive plastic part wall thickness optimization control method are detailed in the above-mentioned embodiment of the adaptive plastic part wall thickness optimization control device, and will not be repeated here.

[0228] Example 4

[0229] This embodiment further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the method described in embodiment 1 are implemented.

[0230] The above embodiments can be implemented in whole or in part via software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product comprises one or more computer instructions or computer programs. When loaded or executed on a computer, the processes or functions described in accordance with the embodiments of the present invention are fully or partially performed. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transferred from one computer-readable storage medium to another. For example, the computer instructions can be transferred from one website, computer, server, or data center to another website, computer, server, or data center via a wired or wireless network. The computer-readable storage medium can be any available medium accessible by a computer, or a data storage device such as a server or data center that contains a collection of one or more available media. The available medium can be magnetic media (e.g., floppy disks, hard disks, or magnetic tapes), optical media (e.g., DVDs), or semiconductor media. The semiconductor media can be a solid-state drive.

[0231] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed in the present invention can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present invention.

[0232] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0233] In the several embodiments provided by the present invention, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units is only one type. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0234] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0235] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0236] The above description is merely illustrative of certain exemplary embodiments of the present invention. It goes without saying that those skilled in the art will be able to modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the above drawings and description are illustrative in nature and should not be construed as limiting the scope of protection of the claims.

Claims

1. An adaptive plastic part wall thickness optimization control method, characterized in that: The following steps are involved: The magnetic field signals at different positions of the target plastic part are collected and processed using wavelet transform to obtain the corresponding wavelet coefficients. The low-dimensional magnetic field feature vector is obtained by dimensionality reduction using the pre-trained first autoencoder model. According to the pre-trained feature association network model, the low-dimensional magnetic field feature vector is cross-mapped with a preset wall thickness state library to obtain the wall thickness state feature and determine the wall thickness fluctuation type; The radial micro-displacement of the mold core is adjusted according to the wall thickness fluctuation type, and the magnetic field signal and mold core position data are collected alternately in real time. The real-time wall thickness response characteristics are obtained through interactive fusion processing. The acquisition logic of the real-time wall thickness response feature is: According to the preset wall thickness fluctuation type mark, the mold core micro-displacement device installed inside the plastic molding mold is triggered to perform radial micro-displacement adjustment; During the adjustment process of the mold core micro-displacement device, the magnetic field signal data detected by the magnetic field sensor arranged along the circumference of the mold outlet and the actual radial displacement data of the mold core detected by the position sensor are collected alternately in real time; Based on the core position data, a data analysis window centered on the current core position is constructed to intercept the corresponding short-term magnetic field signal sequence; The short-term magnetic field signal sequence and the current core position data are simultaneously input into the cross-attention network model for interactive fusion processing, wherein the core position data is used as the query vector, and the magnetic field signal sequence is used as the key vector and value vector to calculate the dynamic interaction weight; Outputting interactive fusion features according to dynamic interactive weights, and performing feature normalization processing on the interactive fusion features to obtain real-time wall thickness response features of the current real-time wall thickness state of the plastic part; The second autoencoder model dynamically reconstructs the difference between the real-time wall thickness response characteristics and the previous wall thickness state characteristics to generate wall thickness trend prediction features; The trend prediction features and the real-time wall thickness response features are input into a reinforcement learning decision network to generate wall thickness optimization control instructions, and the temperature and mold core displacement parameters are controlled in real-time linkage.

2. The adaptive plastic part wall thickness optimization control method according to claim 1, characterized in that: The acquisition logic of the low-dimensional magnetic field characteristic vector is: Acquire the magnetic field signal collected by the magnetic field sensor and process the magnetic field signal according to the wavelet transform method to obtain the corresponding wavelet coefficient; combine the wavelet coefficients of multiple magnetic field sensors to obtain a high-dimensional wavelet coefficient feature vector; The combined processing includes: Select the wavelet coefficients collected by multiple magnetic field sensors at the same time to construct the original wavelet coefficient matrix; The original wavelet coefficient matrix is ​​subjected to standardization preprocessing, specifically: the wavelet coefficients corresponding to each magnetic field sensor in the matrix are respectively subjected to mean zeroing and variance standardization; Perform feature fusion processing on the standardized wavelet coefficient matrix, and use the pre-built autoencoder network to nonlinearly map the standardized wavelet coefficient matrix into a low-dimensional feature representation; For the low-dimensional feature vector after fusion processing, the low-dimensional magnetic field feature vector with the highest sensitivity to the wall thickness change of the plastic part is screened out through the feature selection method to obtain the optimal low-dimensional expression of the high-dimensional wavelet coefficient feature vector.

3. The adaptive plastic part wall thickness optimization control method according to claim 2, characterized in that: The first autoencoder model includes an encoder and a decoder, specifically including: The encoder is used to map the input high-dimensional wavelet coefficient feature vector to a low-dimensional magnetic field feature vector; The decoder is used to reconstruct the low-dimensional magnetic field feature vector back into a high-dimensional wavelet coefficient feature vector; During the training process, the high-dimensional wavelet coefficient eigenvector is used as the model input and training target, and the network parameters are optimized through the error back propagation algorithm to minimize the reconstruction error between the input and the reconstructed eigenvector.

4. The adaptive plastic part wall thickness optimization control method according to claim 3, characterized in that: The specific network structure of the first autoencoder model includes: Input layer, used to receive input high-dimensional wavelet coefficient feature vector; The encoder sequentially sets a multi-layer fully connected network structure with the number of nodes in each layer gradually decreasing, and obtains the low-dimensional magnetic field feature vector in the middle hidden layer; The decoder is symmetrical with the encoder structure. It sets up a multi-layer fully connected network structure in sequence, with the number of nodes in each layer gradually increasing, and outputs the reconstructed low-dimensional magnetic field feature vector.

5. The adaptive plastic part wall thickness optimization control method according to claim 4, characterized in that: The application logic of the feature association network model is: The feature association network model includes an input layer, a hidden layer and an output layer, wherein the input layer receives a low-dimensional magnetic field feature vector, and the output layer outputs an initial feature of the wall thickness state corresponding to the wall thickness state library; By collecting magnetic field signal data of multiple historical plastic part samples, wavelet transform and autoencoder are used to obtain the corresponding low-dimensional magnetic field feature vectors as well as the actual wall thickness state data and wall thickness fluctuation type classification data of multiple historical plastic part samples; The low-dimensional magnetic field feature vectors of the historical plastic part samples and the actual wall thickness state data are combined into a training data set to train the feature association network model; The feature association network model is trained. During training, the low-dimensional magnetic field feature vectors of historical plastic part samples are used as network inputs, and the actual wall thickness fluctuation type is used as the training label. The error back propagation algorithm is used to train the feature association network model. The optimization goal is to minimize the loss between the predicted output and the actual wall thickness fluctuation type. During the training process, the model weights are dynamically adjusted until the classification accuracy of the initial features of the wall thickness state output by the model reaches above the preset accuracy threshold; and the trained feature association network model is output.

6. The adaptive plastic part wall thickness optimization control method according to claim 5, characterized in that: The acquisition logic of the wall thickness trend prediction feature is: Constructing and training a second autoencoder model; the second autoencoder model includes an encoder and a decoder, wherein the encoder is used to perform nonlinear dimensionality reduction processing on the difference features between the real-time wall thickness response features and the initial features of the previous wall thickness state to obtain a low-dimensional difference feature vector; and the decoder is used to nonlinearly reconstruct the low-dimensional difference feature vector into a reconstructed difference feature vector with the same dimension as the input difference feature vector; When training the second autoencoder model, historical wall thickness difference feature samples calculated by the difference between the real-time wall thickness response features at consecutive moments in historical plastic part production and the corresponding initial wall thickness state features are used as input and target. An error backpropagation algorithm is used, with the optimization goal of minimizing the mean square error between the input and the reconstructed output. During the real-time production process, a current difference feature vector between the current real-time wall thickness response feature and the previous wall thickness state initial feature is obtained through differential calculation, and the difference feature vector is input into the pre-trained second autoencoder model. After the encoder dimensionality reduction processing, a low-dimensional difference feature vector is obtained as the wall thickness trend prediction feature; The wall thickness trend prediction feature is matched with a preset historical wall thickness trend feature library, and the wall thickness trend prediction feature is output according to the matching result.

7. The adaptive plastic part wall thickness optimization control method according to claim 6, characterized in that: The application logic of the wall thickness optimization control instruction is: Constructing and pre-training a reinforcement learning decision network, wherein the reinforcement learning decision network includes an input layer, a policy network, and a value network, wherein the input layer simultaneously receives wall thickness trend prediction features and real-time wall thickness response features; The strategy network outputs radial displacement parameters of the mold core micro-displacement device and temperature adjustment parameters of the temperature control device; the value network evaluates the long-term stability of the output action of the strategy network; The reinforcement learning decision network training method comprises: collecting wall thickness trend prediction features and real-time wall thickness response features during the historical plastic part production process as input states, recording the corresponding mold core radial displacement adjustment parameters and temperature adjustment parameters, and the corresponding actual wall thickness change results as feedback rewards; based on the state-action-feedback reward triplet, the network is trained with wall thickness stability as the goal; During the actual plastic part production process, the real-time wall thickness response characteristics and the corresponding wall thickness trend prediction characteristics are simultaneously input into the reinforcement learning decision network to generate wall thickness optimization control instructions in real time. Based on the wall thickness optimization control instructions, the radial displacement parameters of the mold core micro-displacement device and the temperature parameters during the plastic part production process are adjusted in real time to achieve real-time closed-loop linkage control.

8. An adaptive plastic part wall thickness optimization control device, based on the implementation of the adaptive plastic part wall thickness optimization control method according to any one of claims 1 to 7, characterized in that: It includes a magnetic field signal feature extraction module, a wall thickness state recognition module, a real-time wall thickness response feature generation module, a wall thickness trend prediction module and a reinforcement learning decision control module. Data is transmitted between each module via wired and / or wireless communication. The magnetic field signal feature extraction module collects magnetic field signals at different locations of the target plastic part, processes the magnetic field signals at different locations using wavelet transform to obtain corresponding wavelet coefficients, and then reduces the dimension through the pre-trained first autoencoder model to obtain a low-dimensional magnetic field feature vector; A wall thickness state recognition module, which cross-maps the low-dimensional magnetic field feature vector with a preset wall thickness state library based on a pre-trained feature association network model to obtain wall thickness state features and determine the type of wall thickness fluctuation; The real-time wall thickness response feature generation module triggers radial micro-displacement adjustment of the mold core according to the wall thickness fluctuation type, while alternately collecting magnetic field signals and mold core position data in real time, and obtains the real-time wall thickness response feature through interactive fusion processing; The acquisition logic of the real-time wall thickness response feature is: According to the preset wall thickness fluctuation type mark, the mold core micro-displacement device installed inside the plastic molding mold is triggered to perform radial micro-displacement adjustment; During the adjustment process of the mold core micro-displacement device, the magnetic field signal data detected by the magnetic field sensor arranged along the circumference of the mold outlet and the actual radial displacement data of the mold core detected by the position sensor are collected alternately in real time; Based on the core position data, a data analysis window centered on the current core position is constructed to intercept the corresponding short-term magnetic field signal sequence; The short-term magnetic field signal sequence and the current core position data are simultaneously input into the cross-attention network model for interactive fusion processing, wherein the core position data is used as the query vector, and the magnetic field signal sequence is used as the key vector and value vector to calculate the dynamic interaction weight; Outputting interactive fusion features according to dynamic interactive weights, and performing feature normalization processing on the interactive fusion features to obtain real-time wall thickness response features of the current real-time wall thickness state of the plastic part; The wall thickness trend prediction module dynamically reconstructs the difference between the real-time wall thickness response characteristics and the previous wall thickness state characteristics based on the second autoencoder model to generate wall thickness trend prediction features; The reinforcement learning decision control module inputs the trend prediction characteristics and the real-time wall thickness response characteristics into the reinforcement learning decision network, generates wall thickness optimization control instructions, and performs real-time linkage control on the temperature and mold core displacement parameters.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the adaptive plastic part wall thickness optimization control method as claimed in any one of claims 1 to 7 are implemented.

Images