An intelligent temperature regulation method and system for injection molding dies

Abstract

The present application provides an intelligent temperature regulation method and system for an injection molding die. By collecting temperature data and pressure data of each area of the injection molding die, a real-time multi-source sensor data stream is generated; feature extraction is performed on the real-time multi-source sensor data stream to obtain the temperature field distribution characteristics; based on the temperature field distribution characteristics and the differential private Bayesian optimization algorithm, a temperature field distribution prediction model is trained, and then the die temperature prediction result is output through this model; feature fusion is performed on the die temperature prediction result and the pressure data to construct the input features of the neural network controller, and based on the input features and the positive definite curvature learning method, a neural network controller is trained. The power control signal of the heater and the flow control instruction of the water cooling system are generated through the neural network controller, and based on this control signal and instruction, the die temperature field is dynamically controlled through the adjustable gap compensation mechanism. The present application improves the accuracy and response speed of temperature regulation.

Description

Technical Field

[0001] This application relates to the technical field of injection molding die control, and particularly to an intelligent temperature regulation method and system for injection molding dies. Background Art

[0002] Injection molding is an important industrial production process, and the accuracy and stability of die temperature control directly affect the quality and production efficiency of products. With the continuous improvement of industrial production requirements for product quality, the precise control of the die temperature field has become a key technical difficulty.

[0003] Currently, die temperature control mainly adopts traditional control methods such as PID control, and adjusts the heater power and the flow rate of the cooling system by setting fixed control parameters. Although this method is simple to implement, due to the lack of accurate modeling and prediction capabilities for the temperature field distribution characteristics, it is difficult to adapt to the temperature dynamic changes under complex working conditions.

[0004] In recent years, with the development of artificial intelligence technology, temperature control methods based on machine learning have gradually been applied to die temperature control. These methods have achieved a certain degree of intelligent control by establishing temperature prediction models and control models. However, the existing technologies have problems such as low temperature control accuracy, difficulty in effectively coping with dynamic changes during the molding process, and slow response speed that is difficult to meet the real-time control requirements. Summary of the Invention

[0005] In view of this, this application provides an intelligent temperature regulation method and system for injection molding dies, which solves the problems in the prior art of low dynamic control accuracy of die temperature, difficulty in effectively coping with dynamic changes during the molding process, and slow response speed that is difficult to meet the real-time control requirements.

[0006] An embodiment of this application provides an intelligent temperature regulation method for injection molding dies, including:

[0007] Collect temperature data and pressure data of each area of the injection molding die to generate a real-time multi-source sensor data stream;

[0008] Extract features from the real-time multi-source sensor data stream through the Transformer model and the EM algorithm to obtain the temperature field distribution characteristics;

[0009] Based on the temperature field distribution characteristics and the differential private Bayesian optimization algorithm, train to obtain a temperature field distribution prediction model, and output the die temperature prediction result through the temperature field distribution prediction model;

[0010] Fuse the characteristics of the predicted results of the mold temperature and the pressure data to construct the input characteristics of the neural network controller, train the neural network controller based on the input characteristics and the positive definite curvature learning method, and generate the power control signal of the heaters in each area of the injection molding die and the flow control instruction of the water cooling system through the neural network controller;

[0011] Based on the power control signal of the heaters and the flow control instruction of the water cooling system, dynamically control the temperature field of the injection molding die through an adjustable gap compensation mechanism.

[0012] Optionally, collecting the temperature data and pressure data of each area of the injection molding die to generate a real-time multi-source sensor data stream, including:

[0013] Through a high-speed data acquisition card, perform analog-to-digital conversion on the temperature data collected by the temperature sensor and the pressure data collected by the pressure sensor to obtain digital signals;

[0014] Perform signal filtering, data calibration, and outlier removal on the digital signals to generate preprocessed data;

[0015] Establish a distributed data caching mechanism to synchronize and align the preprocessed data according to timestamps to generate a real-time multi-source sensor data stream.

[0016] Optionally, extracting the characteristics of the real-time multi-source sensor data stream through the Transformer model and the EM algorithm to obtain the temperature field distribution characteristics, including:

[0017] Add temporal and spatial position information to the real-time multi-source sensor data stream through the position encoder of the Transformer model to generate an initial feature representation;

[0018] Based on the initial feature representation, calculate the similarity between the query matrix, the key matrix, and the value matrix using the multi-head attention mechanism to generate an attention weight matrix reflecting the spatio-temporal correlation of the data;

[0019] Adopt a residual connection structure to perform weighted fusion of the attention weight matrix and the initial feature representation to obtain an enhanced feature representation;

[0020] Perform clustering analysis on the enhanced feature representation through the EM algorithm to obtain the temperature field distribution characteristics.

[0021] Optionally, performing clustering analysis on the enhanced feature representation through the EM algorithm to obtain the temperature field distribution characteristics, including:

[0022] Based on the enhanced feature representation, randomly initialize K clustering centers to obtain initial clustering parameters, where the value of K is adaptively determined by the silhouette coefficient;

[0023] According to the initial clustering parameters, alternately execute the steps of calculating the data point membership probability and updating the clustering centers, and dynamically adjust the parameter update step size through an adaptive learning rate mechanism until the target clustering result is obtained;

[0024] Through the target clustering result, obtain the temperature field distribution characteristics, where the temperature field distribution characteristics include a temperature gradient vector field, a hot spot distribution feature map, and a temperature fluctuation spectrum.

[0025] Optionally, training to obtain a temperature field distribution prediction model based on the temperature field distribution characteristics and the differential privacy Bayesian optimization algorithm includes:

[0026] Based on the temperature field distribution characteristics, construct the prior distribution of the Bayesian probability model;

[0027] Randomly perturb the model parameters through the differential privacy mechanism, and search for the target model parameters through the Bayesian optimization method;

[0028] Based on the target model parameters, train to obtain a temperature field distribution prediction model.

[0029] Optionally, constructing the prior distribution of the Bayesian probability model based on the temperature field distribution characteristics includes:

[0030] Through multi-dimensional Gaussian distribution modeling and historical data analysis, obtain the initial prior distribution based on the temperature gradient vector field and the hot spot distribution feature map in the temperature field distribution characteristics;

[0031] Based on the initial prior distribution, automatically learn the parameter distribution of different working conditions through a hierarchical Bayesian structure to generate a hierarchical prior model;

[0032] Based on the hierarchical prior model, construct a Bayesian probability model and output the prior distribution parameters;

[0033] And, randomly perturbing the model parameters through the differential privacy mechanism and searching for the target model parameters through the Bayesian optimization method includes:

[0034] Add random noise to the prior distribution parameters through the Laplace mechanism, and dynamically adjust the privacy budget based on the parameter sensitivity to obtain the perturbed parameters;

[0035] Construct a surrogate model through Gaussian process regression to evaluate the parameter performance of the perturbed parameters, and obtain a candidate parameter set through an exploration and exploitation balance strategy, where the evaluation indicators of the surrogate model include prediction accuracy and computational efficiency;

[0036] For the candidate parameter set, the model parameters are optimized by variational inference method and minimizing the evidence lower bound method, and the target model parameters are output.

[0037] Optionally, the feature fusion of the die temperature prediction result and the pressure data to construct the input features of the neural network controller includes:

[0038] Through a sliding time window mechanism, the die temperature prediction result and the real-time pressure data are constructed into a time series feature sequence;

[0039] Through an attention mechanism, the time series feature sequence is weighted to obtain a weighted feature sequence;

[0040] The weighted feature sequence is subjected to feature vector transformation to generate the input features of the neural network controller.

[0041] Optionally, the method for training the neural network controller based on the input features and the positive definite curvature learning method includes:

[0042] According to the input features, a loss function based on Riemannian metric is constructed, where the loss function includes a temperature control error term, a pressure balance term, and an energy consumption constraint term, and a curvature regularization term is obtained to obtain an optimization objective function;

[0043] According to the optimization objective function, the Riemannian gradient is calculated in the tangent space of the parameters, and the updated parameters are projected back to the manifold space satisfying the positive definite constraint through exponential mapping to obtain the optimized network parameters.

[0044] Optionally, the dynamic control of the temperature field of the injection molding die by the adjustable gap compensation mechanism based on the power control signal of the heater and the flow control instruction of the water cooling system includes:

[0045] Based on the power control signal of the heater, the power of the heaters in each area of the injection molding die is modulated by a thyristor module, and based on the flow control instruction of the water cooling system, the cooling water flow of the water cooling system in each area of the injection molding die is controlled by a high-precision proportional valve;

[0046] The displacement change of the adjustable gap compensation mechanism is monitored in real time, and the gap size of the cavity of the injection molding die is adjusted based on the displacement change;

[0047] Based on the temperature data and pressure data of the injection molding die collected in real time, the PID parameters are adjusted through a fuzzy adaptive algorithm, and the prediction period and control time domain of the model predictive controller are optimized based on the reinforcement learning method to obtain optimized control parameters. When it is detected that the process parameters meet the preset change conditions, a parameter re-optimization process is triggered.

[0048] An embodiment of the present application also provides an intelligent temperature regulation system for an injection molding die, including:

[0049] A data acquisition module for collecting temperature data and pressure data of each area of the injection molding die to generate a real-time multi-source sensor data stream;

[0050] A feature recognition module for extracting features from the real-time multi-source sensor data stream through a Transformer model and an EM algorithm to obtain temperature field distribution features;

[0051] A temperature prediction module for training a temperature field distribution prediction model based on the temperature field distribution features and the differential private Bayesian optimization algorithm, and outputting a die temperature prediction result through the temperature field distribution prediction model;

[0052] A control strategy module for performing feature fusion on the die temperature prediction result and the pressure data to construct input features of a neural network controller, training the neural network controller based on the input features and the positive definite curvature learning method, and generating power control signals for heaters in each area of the injection molding die and flow control instructions for a water cooling system through the neural network controller;

[0053] An execution control module for dynamically controlling the temperature field of the injection molding die through an adjustable gap compensation mechanism based on the power control signals of the heaters and the flow control instructions of the water cooling system.

[0054] The present application has the following technical effects:

[0055] This application collects temperature data and pressure data in each area of an injection molding die to generate a real-time multi-source sensor data stream; through a Transformer model and an EM algorithm, feature extraction is performed on the real-time multi-source sensor data stream to obtain temperature field distribution features; based on the temperature field distribution features and a differential privacy Bayesian optimization algorithm, a temperature field distribution prediction model is trained, and the die temperature prediction result is output through the temperature field distribution prediction model; feature fusion is performed on the die temperature prediction result and the pressure data to construct the input features of a neural network controller, and based on the input features and a positive definite curvature learning method, a neural network controller is trained, and through the neural network controller, power control signals for heaters in each area of the injection molding die and flow control instructions for a water cooling system are generated; based on the power control signals for heaters and the flow control instructions for the water cooling system, the temperature field of the injection molding die is dynamically controlled through an adjustable gap compensation mechanism. Among them, feature extraction of multi-source sensor data through a Transformer model and an EM algorithm realizes high-precision identification of temperature field distribution features, a differential privacy Bayesian optimization algorithm is used to train the temperature field distribution prediction model, improving the model prediction accuracy while protecting data privacy, and a neural network controller is trained based on a positive definite curvature learning method, enhancing the stability and convergence of the controller, and the dynamic control of the temperature field is realized through an adjustable gap compensation mechanism, significantly improving the accuracy and response speed of temperature regulation. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] To more clearly illustrate the technical solutions of the embodiments of the present disclosure, the drawings required for use in the embodiments will be briefly introduced below. The drawings here are incorporated into the specification and constitute a part of this specification. These drawings show embodiments that conform to the present disclosure and are used together with the specification to illustrate the technical solutions of the present disclosure. It should be understood that the following drawings only show certain embodiments of the present disclosure and should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can also be obtained based on these drawings without creative efforts.

[0057] Figure 1 It is a schematic flowchart of an intelligent temperature regulation method for an injection molding die provided by an embodiment of this application;

[0058] Figure 2 It is a schematic flowchart of a method for generating a real-time multi-source sensor data stream in an embodiment of this application;

[0059] Figure 3 It is a schematic structural diagram of a feature extraction method based on a Transformer model and an EM algorithm in an embodiment of this application;

[0060] Figure 4Schematic diagram of the training process of the temperature field distribution prediction model in the embodiments of the present application;

[0061] Figure 5 Schematic diagram of the structure of the neural network controller in the embodiments of the present application;

[0062] Figure 6 Schematic diagram of the structure of the intelligent temperature regulation system for injection molding dies in the embodiments of the application. Detailed implementation manners

[0063] To make the objectives, technical solutions, and advantages of the embodiments of the present disclosure clearer, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present disclosure. Apparently, the described embodiments are only a part rather than all of the embodiments of the present disclosure. Components of the embodiments of the present disclosure generally described and illustrated in the figures herein may be arranged and designed in a variety of different configurations. Therefore, the detailed description of the embodiments of the present disclosure provided in the figures is not intended to limit the scope of the claimed present disclosure, but is merely representative of selected embodiments of the present disclosure. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present disclosure without creative efforts fall within the scope of the present disclosure.

[0064] Figure 1 Schematic diagram of the process of the intelligent temperature regulation method for injection molding dies provided in the embodiments of the present application. As Figure 1 shown, the embodiments of the present application provide an intelligent temperature regulation method for injection molding dies, including the following steps S1 to S5:

[0065] S1: Collect the temperature data and pressure data of each area of the injection molding die to generate a real-time multi-source sensor data stream;

[0066] As Figure 2 shown, in some embodiments, step S1 may specifically include steps S1.1 to S1.3:

[0067] S1.1: Through a high-speed data acquisition card, perform analog-to-digital conversion on the temperature data collected by the temperature sensor and the pressure data collected by the pressure sensor to obtain digital signals;

[0068] S1.2: Perform signal filtering, data calibration, and outlier rejection on the digital signals to generate preprocessed data;

[0069] S1.3: Establish a distributed data caching mechanism to synchronize and align the preprocessed data according to timestamps to generate a real-time multi-source sensor data stream.

[0070] In this embodiment, temperature sensor arrays (including thermocouples, infrared sensors, etc.) and pressure sensors are arranged at key positions of the mold to collect temperature data and pressure data on the surface and inside of the mold in real time. Among them, the sampling frequency of the temperature sensor is set to 100 Hz, and the sampling frequency of the pressure sensor is set to 200 Hz. Then, a high-speed data acquisition card is used to perform A / D conversion and preprocessing on the original sensor signals collected. Specifically, it includes signal filtering, data calibration, and outlier rejection to ensure the accuracy and reliability of the data. Next, a distributed data caching mechanism is established to synchronize and align the preprocessed temperature and pressure data according to the timestamps to form a unified data format. Finally, a real-time data stream processing pipeline is constructed to package the synchronized multi-source sensor data into a unified data stream format to achieve real-time data transmission and storage.

[0071] Specifically, during the intelligent temperature regulation process of the injection molding die, due to differences in factors such as the sampling frequency, response time, and data transmission delay of the temperature sensor and the pressure sensor, the data collected by different sensors often show asynchronous phenomena in time. To solve this problem, this application uses a distributed data caching mechanism to synchronize and align multi-source sensor data.

[0072] Specifically, first, multiple levels of data buffer areas are set in the system, including local caches and distributed caches. The local cache is used to temporarily store the original data of each sensor, and the distributed cache uses a Redis cluster to store the preprocessed data. The system adds a unified timestamp identifier to each sensor data packet, which is generated based on a high-precision clock source and is accurate to the microsecond level to ensure the accuracy of the time mark.

[0073] During the data synchronization process, the system sets a sliding time window with a fixed size, and the window size is dynamically adjusted according to the real-time requirements of the system and data characteristics, usually set to 10 - 100 ms. For different sensor data falling within the same time window, the system first caches them into the corresponding local buffer area. When the data volume in the local buffer area reaches the preset threshold, the data synchronization mechanism is triggered.

[0074] The data alignment process is carried out in a time-driven manner. The system uses the timestamp of the temperature data as a reference to find the closest pressure data within the corresponding time window. If it is found that there is a lack of corresponding sensor data in a certain time window, data compensation is performed through a linear interpolation algorithm. To ensure the timeliness of the data, the system will set a data waiting timeout threshold, and data exceeding this threshold will be discarded.

[0075] In addition, to handle anomalies such as network latency and data loss, the system implements a data retransmission and recovery mechanism. When a data packet loss is detected, the system requests retransmission from the data source and uses data caching for temporary compensation to ensure the continuity of the data stream. The system also monitors the working status of each sensor in real time through a heartbeat detection mechanism and performs fault handling in a timely manner when a sensor anomaly is found.

[0076] Finally, the synchronized and aligned data is organized according to a unified time series to form a real-time multi-source sensor data stream in a standard format. Each data point in this data stream contains temperature and pressure information at the corresponding moment, providing high-quality input data for subsequent feature extraction and model training.

[0077] S2: Extract features from the real-time multi-source sensor data stream through a Transformer model and an EM algorithm to obtain temperature field distribution features;

[0078] As Figure 3 shown, in some embodiments, step S2 may specifically include steps S2.1 to S2.4:

[0079] S2.1: Add temporal position information to the real-time multi-source sensor data stream through the position encoder of the Transformer model to generate an initial feature representation.

[0080] In this embodiment, when performing temporal feature extraction, the system receives a multi-source sensor data stream as input. Specifically, the improved Transformer encoder adopts a multi-layer stacked structure, and each layer contains a self-attention module and a feed-forward neural network. For the input temperature data sequence {T1, T2,..., Tn} and pressure data sequence {P1, P2,..., Pn}, temporal position information is added through the position encoder to generate an initial feature representation. This improved encoder structure can not only capture the long-term dependencies of the data but also maintain the integrity of the temporal information.

[0081] Specifically, the system first processes the real-time multi-source sensor data stream using the position encoder of the Transformer model. The purpose of position encoding is to assign relative position information to each temporal data point so that the model can perceive the temporal relationship of the data. Specifically, the system uses a combination of sine and cosine functions to generate position encoding vectors, and the dimension of the position encoding vector is the same as the data feature dimension.

[0082] For time step t and position i, the calculation formula for the position encoding PE(t,i) is:

[0083]

[0084] where d is the feature dimension. By adding the positional encoding vector to the original data feature vector, an initial feature representation containing temporal and spatial position information is obtained.

[0085] S2.2: Based on the initial feature representation, use the multi-head attention mechanism to calculate the similarity between the query matrix, the key matrix, and the value matrix, and generate an attention weight matrix for reflecting the spatio-temporal correlation of the data.

[0086] In this embodiment, on the basis of feature extraction, the system designs an innovative multi-head attention mechanism. Specifically, the number of attention heads is set to 8, and each attention head is responsible for capturing feature patterns at different scales. By calculating the similarity between the query matrix (Query), the key matrix (Key), and the value matrix (Value), the system can automatically learn the association relationships between different sensor data. Specifically, the outputs of multiple attention heads are concatenated and linearly transformed to obtain a complete attention weight matrix. For example, when the temperature in a certain area rises abnormally, the system will automatically pay attention to the temperature changes and pressure changes in adjacent areas, thereby generating an attention weight matrix reflecting the spatio-temporal correlation of the data. The multi-head attention mechanism is implemented to capture the spatio-temporal correlation between data. Each of the 8 attention heads focuses on different feature subspaces, so as to comprehensively capture the multi-scale correlation of the data. S2.3: Adopt a residual connection structure to perform weighted fusion of the attention weight matrix and the initial feature representation to obtain an enhanced feature representation.

[0087] In this embodiment, a residual connection structure is adopted to perform weighted fusion of the attention weight matrix and the initial feature representation. Specifically, during implementation, the attention weights are first normalized to ensure that the sum of the weights is 1, and then weighted summation is performed with the initial feature representation. In addition, a layer normalization mechanism is introduced, which effectively alleviates the problem of gradient disappearance in the training of deep networks and retains the original feature information, improving the stability of feature extraction. Specifically, during implementation, the system performs element-wise addition of the attention weight matrix and the initial feature representation, and then undergoes layer normalization processing to obtain an enhanced feature representation.

[0088] Layer normalization helps to stabilize the training process and accelerate convergence, and its calculation formula is:

[0089]

[0090] where and are the feature mean and standard deviation respectively, and are learnable scaling and bias parameters.

[0091] S2.4: Perform clustering analysis on the enhanced feature representation through the EM algorithm to obtain the temperature field distribution characteristics.

[0092] Among them, step S2.4 may specifically include:

[0093] S2.4.1: Based on the enhanced feature representation, randomly initialize K clustering centers to obtain initial clustering parameters, where the value of K is adaptively determined by the silhouette coefficient;

[0094] S2.4.2: According to the initial clustering parameters, alternately execute the steps of calculating the belonging probability of data points and updating the clustering centers, and dynamically adjust the parameter update step size through the adaptive learning rate mechanism until the target clustering result is obtained;

[0095] S2.4.3: Obtain the temperature field distribution characteristics through the target clustering result, where the temperature field distribution characteristics include a temperature gradient vector field, a hot spot distribution feature map, and a temperature fluctuation spectrum.

[0096] In this embodiment, the improved EM algorithm is applied to feature clustering analysis. This algorithm first adaptively determines the optimal number of clusters K through the silhouette coefficient and randomly initializes K clustering centers. The silhouette coefficient is an index to measure the clustering effect, and its value ranges from -1 to 1. The closer it is to 1, the better the clustering effect. The system will automatically test different values of K (generally between 3 and 10), calculate the silhouette coefficient corresponding to each value of K, and select the K value with the highest score. For example, assume that the silhouette coefficient is 0.65 when testing K = 3, 0.78 when K = 4, and 0.72 when K = 5. Then finally select K = 4 as the number of clusters. After determining the value of K, use the improved initialization method to select the clustering centers. This method will preferentially select points farther away from the existing clustering centers as the new clustering centers to avoid the initial clustering centers being too concentrated.

[0097] After determining the value of K, the EM algorithm alternately executes the expectation step (E-step) and the maximization step (M-step). In the E-step, calculate the posterior probability that each data point belongs to each cluster. In the M-step, update the clustering centers and covariance matrices based on the posterior probability. And to improve the robustness of the algorithm, an adaptive learning rate mechanism is introduced to dynamically adjust the parameter update step size until the clustering centers converge to a stable state. For example, if the initial learning rate is set to 0.01, as the number of iterations increases, the learning rate will gradually decrease. When the algorithm runs to the 100th iteration, the learning rate may drop to 0.001, and to the 500th iteration may drop to 0.0001. The system will monitor the parameter changes. If the parameter changes in 10 consecutive iterations are all less than 0.000001, it is considered that the algorithm has converged and the iteration can be stopped.

[0098] Through the iterative optimization of the EM algorithm, the system finally obtains the temperature field distribution characteristics, that is, a series of key characteristics characterizing the temperature field distribution, specifically including:

[0099] The temperature gradient vector field is constructed by calculating the temperature difference between adjacent measurement points and is used to describe the direction and intensity of heat transfer. For example, if the temperature at a certain point is 150°C, the temperature of the point above it is 148°C, and the temperature of the point on the right is 152°C, then the heat transfer trend at this point is to diffuse upward and receive heat from the right;

[0100] The hot spot distribution feature map reflects the position and range of temperature anomaly regions. First, calculate the average value and fluctuation range of the overall temperature, and then mark the regions that exceed the normal range. For example, if the temperature in a certain region continuously exceeds the average temperature by more than 20°C, it will be marked as a hot spot region;

[0101] The temperature fluctuation spectrum characterizes the periodic characteristics of temperature changes, etc. Specifically, analyze the change characteristics of temperature data on different time scales, and identify periodic temperature fluctuation patterns. For example, the system may find that the temperature fluctuates once every 30 seconds in some regions, which may be related to the injection molding cycle; or find that there are high-frequency temperature jitters in some regions, which may indicate unstable heater control. The system focuses on the temperature changes in the time scale from 0.1 second to 10 seconds because this range usually contains the most important process-related fluctuations.

[0102] These multi-dimensional feature information provides an important basis for subsequent temperature field prediction and optimization of control strategies. It should be noted that the system uses the feature importance analysis method to only retain the features that make a significant contribution to the characterization of the temperature field, which not only improves the calculation efficiency but also reduces the complexity of subsequent modeling.

[0103] In addition, in this embodiment, the Transformer model mainly processes the time series data of the mold temperature field. Taking an actual scenario as an example, assume that there are 12 temperature sensors and 8 pressure sensors arranged on the mold, the sampling frequency is 10Hz, and the system processes a 30-second data window each time. At this time, the input of each time step is a 20-dimensional vector (including 12 temperature values and 8 pressure values), and the entire sequence contains 300 time steps. The input shape of the model is [batch size, 300, 20], where the batch size is usually set to 32 or 64. These data will be normalized and filled with missing values before being input into the model.

[0104] Positional encoding is an important part of the Transformer model. The system generates a positional encoding vector with the same dimension as the feature dimension for each time step, which is 20 in this example. These encoding vectors help the model understand the relative positional relationships of data points in the time series, enabling the model to distinguish data at different time steps. The positional encoding vectors are generated through trigonometric functions and can express the absolute position and relative distance information of data points.

[0105] The multi-head attention mechanism is the core of the model. This system uses 8 attention heads, each with a dimension of 64. The original 20-dimensional input features are first mapped to a 512-dimensional space and then divided into 8 64-dimensional subspaces. Each attention head independently calculates the attention weights in its corresponding subspace. This design allows the model to simultaneously focus on different types of feature patterns. For example, some attention heads may focus on the long-term change trend of temperature, while others focus on instantaneous fluctuations.

[0106] The features extracted by the model can be divided into three categories: The first category is time-series dependent features, which describe the change patterns of temperature and pressure over time. For example, the readings of a certain temperature sensor show a peak every 5 seconds, which is usually related to the injection molding cycle. The second category is spatial correlation features, which describe the relationships between different sensors. For example, after the readings of pressure sensor P1 increase, the readings of temperature sensors T3 and T4 will increase 2 seconds later. The third category is multi-scale features, which are features at various time scales captured by different attention heads, including fast temperature fluctuations at the millisecond level and temperature gradual change trends at the second level.

[0107] The output of the model is a high-dimensional feature tensor with dimensions [batch size, 300, feature dimension], and the feature dimension is usually set to 256 or 512. These features are then fed into the EM algorithm for clustering analysis to identify the distribution patterns of the temperature field. The output features not only retain the time-series information of the original data but also extract the deep associations between the data, providing a basis for the system to understand and predict the change laws of the temperature field.

[0108] In practical applications, these features exhibit good predictive capabilities. For example, the system can predict possible temperature anomalies in a certain area of the mold in the next few seconds based on these features, or identify the lag relationships between certain sensor data. This information is of great value for optimizing the mold temperature control strategy. In addition, by analyzing the weight distributions of different attention heads, the system can also identify the key factors that have the greatest impact on the temperature field changes, which provides an important basis for process parameter optimization.

[0109] It should be noted that the combination of the improved Transformer model and the EM algorithm not only improves the accuracy of feature extraction, but also significantly improves the real-time performance of the algorithm, meeting the strict requirements for real-time control in the injection molding process.

[0110] S3: Based on the temperature field distribution characteristics and the differential private Bayesian optimization algorithm, train to obtain a temperature field distribution prediction model, and output the die temperature prediction result through the temperature field distribution prediction model;

[0111] As Figure 4 shown, step S3 may specifically include S3.1 to S3.3:

[0112] S3.1: Based on the temperature field distribution characteristics, construct the prior distribution of the Bayesian probability model;

[0113] Step S3.1 may specifically include:

[0114] S3.1.1: Through multi-dimensional Gaussian distribution modeling and historical data analysis, obtain the initial prior distribution based on the temperature gradient vector field and the hot spot distribution feature map in the temperature field distribution characteristics;

[0115] S3.1.2: Based on the initial prior distribution, automatically learn the parameter distribution of different working conditions through a hierarchical Bayesian structure to generate a hierarchical prior model;

[0116] S3.1.3: Based on the hierarchical prior model, construct a Bayesian probability model and output the prior distribution parameters.

[0117] In this embodiment, the prior distribution of the Bayesian probability model is constructed based on the temperature field distribution characteristics obtained in the previous steps. For features such as temperature gradient and hot spot distribution, a multi-dimensional Gaussian distribution is used as the prior model. By analyzing the variation law of the temperature field in historical data, the mean vector and covariance matrix of the prior distribution are determined. The mean vector of the distribution is determined by the average temperature gradient in historical data, and the covariance matrix reflects the correlation of temperature gradients in different regions. For the hot spot distribution feature, the system uses a Bernoulli distribution to describe the probability of each region becoming a hot spot, and the probability parameter is estimated based on the frequency of hot spots appearing in historical data. The temperature fluctuation spectrum feature is modeled by a mixture Gaussian distribution, and each Gaussian component corresponds to a typical temperature fluctuation mode. The selection of these prior distributions fully considers the physical characteristics of the temperature field and can effectively capture the uncertainty of the temperature distribution. In addition, to enhance the adaptability of the model, a hierarchical Bayesian structure is introduced, enabling the model to automatically learn the parameter distribution characteristics under different working conditions.

[0118] Specifically, assume that an injection mold has 16 temperature monitoring points, forming a 4×4 grid layout. In addition to the temperature value, each monitoring point also contains temperature gradient information (in the X and Y directions). Historical data shows that there are three typical operating conditions for this mold during normal production: the standard condition (mold temperature around 160°C), the high-temperature condition (around 180°C), and the low-temperature condition (around 140°C).

[0119] In S3.1.1, the system first processes the temperature gradient vector field data. For the standard condition, the average value of the temperature gradient is smaller in the central region of the mold (about 0.2°C / cm) and larger in the edge region (about 0.5°C / cm), which reflects the characteristic of heat diffusion from the heating point to the surroundings. The hot spot distribution characteristic shows that in the standard condition, hot spots are likely to form in the upper right and lower left regions of the mold, with occurrence probabilities of 15% and 12% respectively. The system uses a 24-dimensional Gaussian distribution (16 temperature values + 8 gradient values) to describe these characteristics, where the mean vector reflects the typical values and the covariance matrix captures the correlations between the characteristics.

[0120] In S3.1.2, the system constructs a three-layer Bayesian structure: the first layer is the operating condition level, which describes the probability distribution of different operating conditions; the second layer is the region level, which describes the temperature characteristic distribution of each region under a specific operating condition; the third layer is the time level, which describes the characteristics of temperature evolution over time. Taking the high-temperature condition as an example, the system finds that the standard deviation of the temperature fluctuation (about 1.5°C) in the upper right region of the mold under this condition is significantly higher than that in the standard condition (about 0.8°C), and this information is encoded into the conditional probability distribution of this region.

[0121] In S3.1.3, the system outputs the complete prior distribution parameters based on the hierarchical structure. Taking the standard condition as an example:

[0122] Overall distribution of the temperature field: mean 160°C, standard deviation 2°C;

[0123] Regional characteristics: average temperature gradient in the central region is 0.2°C / cm, and in the edge region is 0.5°C / cm;

[0124] Hot spot probability: 0.15 in the upper right corner, 0.12 in the lower left corner, and below 0.05 in other regions;

[0125] Time characteristics: temperature fluctuation period is about 30 seconds, corresponding to the injection cycle.

[0126] For the operating condition conversion, the system finds that it takes about 90 seconds to heat up from the standard condition to the high-temperature condition, and the upper right region usually reaches the target temperature first during this period. Such time series characteristics are encoded into the conversion probability matrix. The system also identifies that when the mold cools down from the high-temperature condition, the cooling rate in the edge region (about 0.4°C / second) is faster than that in the central region (about 0.2°C / second).

[0127] In practical applications, this prior knowledge helps the system quickly identify the current working conditions and predict their evolution trends. For example, when the system observes that the temperature in the upper right corner area starts to fluctuate and increase, it can anticipate in advance that it may be transitioning to a high-temperature working condition. Another example is that when it observes that the temperature gradient in the central area increases abnormally (exceeding 0.4 °C / cm), the system will give an early warning of a possible heater failure.

[0128] This hierarchical prior structure provides reliable initial conditions for subsequent predictive modeling. The system can quickly locate the current most likely working conditions and development trends based on the characteristics of the real-time observed temperature field, thereby achieving more accurate predictions. At the same time, this structure also facilitates the online update of the model: when new data is collected, the system can automatically update the parameter distribution of the corresponding layer without having to reconstruct the entire model.

[0129] An important advantage of this hierarchical Bayesian structure is its strong interpretability. When the system makes a certain prediction, it is possible to clearly trace which prior knowledge led to this prediction, which is very helpful for decision-making support and fault diagnosis in actual production. For example, if the system predicts that overheating may occur in a certain area, the basis for this prediction can be understood by checking the historical hot spot probability and the current temperature gradient in that area.

[0130] S3.2: Randomly perturb the model parameters through the differential privacy mechanism, and search for the target model parameters through the Bayesian optimization method;

[0131] S3.2 specifically includes:

[0132] S3.2.1: Add random noise to the prior distribution parameters through the Laplace mechanism, and dynamically adjust the privacy budget based on parameter sensitivity to obtain the perturbed parameters;

[0133] S3.2.2: Construct a surrogate model through Gaussian process regression, evaluate the parameter performance of the perturbed parameters, and obtain a candidate parameter set through an exploration and exploitation balance strategy, where the evaluation metrics of the surrogate model include prediction accuracy and computational efficiency;

[0134] S3.2.3: Optimize the model parameters for the candidate parameter set through the variational inference method and the method of minimizing the evidence lower bound, and output the target model parameters.

[0135] In this embodiment, a differential privacy mechanism is introduced to randomly perturb the model parameters, and the Bayesian optimization method is used to search for the optimal parameters. Specifically, the Laplace mechanism is used to add random noise to the sensitive parameters, and the magnitude of the noise is controlled by the privacy budget ε. A smaller ε value provides stronger privacy protection but may affect the model performance. The Bayesian optimization process uses a Gaussian process as a surrogate model and selects the next set of parameters to be evaluated according to the expected improvement criterion. The optimization goal is to maximize the prediction accuracy of the model while considering the computational efficiency and the privacy protection level. At the same time, an adaptive privacy allocation strategy is designed to dynamically adjust the privacy budget according to the sensitivity of different parameters, maximizing the model performance while protecting data privacy. For example, for the parameters in the area with drastic temperature changes, the system will allocate a larger privacy budget to ensure the prediction accuracy.

[0136] In addition, the system also adopts an adaptive sampling strategy based on Thompson sampling. This strategy dynamically balances the relationship between exploration and exploitation by constructing the posterior distribution of the parameters. In each iteration, a candidate parameter set is sampled from the posterior distribution, and the expected benefits of these parameters are evaluated through the surrogate model. It should be noted that the system uses Gaussian process regression as the surrogate model, which can effectively estimate the objective function value in the unknown parameter space and greatly improve the efficiency of parameter search.

[0137] In S3.2.1, the system first evaluates the sensitivity of each parameter. Assume that the model contains key parameters such as the mean value, variance, and hot spot probability of the temperature field. The sensitivity of the temperature mean is relatively low (changing by 1°C has little impact on the prediction), while the sensitivity of the hot spot probability is relatively high (changing by 0.1 may cause a significant change in the prediction result). Based on this analysis, the system adopts a dynamic privacy budget allocation strategy: allocate a smaller privacy budget (ε = 0.1) to the temperature mean parameter, allowing a larger noise to be added; allocate a larger privacy budget (ε = 0.5) to the hot spot probability to maintain a higher accuracy. For example, the original hot spot probability in a certain area is 0.15, and the temperature mean is 165°C. After adding noise, it may become a probability of 0.16 and a temperature of 163.5°C. This differential privacy protection strategy protects sensitive information while maintaining the basic prediction ability of the model.

[0138] In S3.2.2, the system constructs a Gaussian process surrogate model to evaluate the parameter performance. Taking temperature prediction as an example, the inputs of the surrogate model include the perturbed parameter sets (such as temperature mean, variance, hot spot probability, etc.), and the outputs include the prediction accuracy (average prediction error) and the calculation time. The system adopts a batch sampling strategy and selects 10 groups of candidate parameters for evaluation in each round. In the balance between exploration and exploitation, the system is more inclined to explore new parameter combinations in the early stage (such as the first 30% of the evaluation rounds), and then searches more around the known high-performance parameter regions in the later stage. For example, if it is found that the parameter combinations with the hot spot probability in the range of 0.14 - 0.17 generally perform well, detailed searches will be focused on this range in the later stage.

[0139] In S3.2.3, the system performs the final optimization on the candidate parameters. Suppose the system identifies three groups of candidate parameter configurations: Configuration A focuses on short-term prediction accuracy, Configuration B performs excellently in computational efficiency, and Configuration C strikes a balance between the two. The system approximates the posterior distribution through variational inference and optimizes by minimizing the evidence lower bound (ELBO). The optimization process may find that: Although Configuration C is not the best in individual metrics, its overall performance is the best. For example, the prediction error of Configuration C is only 0.2 °C higher than that of A, but the calculation speed is 40% faster than A, and the privacy protection level is comparable.

[0140] An example of a specific optimization process:

[0141] Initial state:

[0142] Temperature mean parameter: 165 °C (163.5 °C after perturbation);

[0143] Temperature variance: 2.5 °C (2.8 °C after perturbation);

[0144] Hot spot probability: 0.15 (0.16 after perturbation);

[0145] Prediction error: ±1.5 °C;

[0146] Calculation time: 200 ms / prediction;

[0147] Intermediate optimization stage:

[0148] Multiple groups of parameter configurations are tried and it is found that:

[0149] Improving the temperature mean accuracy (reducing the perturbation) can reduce the prediction error to ±1.2 °C, but the calculation time increases to 250 ms;

[0150] Reducing the hot spot probability accuracy can reduce the calculation time to 150 ms, but the prediction error increases to ±1.8 °C;

[0151] Adjusting the variance parameter has relatively little impact on the performance;

[0152] Final optimization result:

[0153] The system selected a balanced configuration:

[0154] Average temperature: 164.2 °C (moderate perturbation);

[0155] Temperature variance: 2.7 °C (large perturbation);

[0156] Hot spot probability: 0.157 (small perturbation);

[0157] Prediction error: ±1.3 °C;

[0158] Computation time: 180 ms / prediction;

[0159] This final configuration performs well in practical applications. For example, when the mold temperature changes rapidly, the system can predict the temperature distribution in the next 5 seconds within 200 ms, with the prediction error controlled within ±1.3 °C. At the same time, through differential privacy protection, even if someone obtains the model parameters, they cannot accurately infer the original temperature distribution characteristics.

[0160] In production practice, the advantage of this optimization strategy lies in its strong adaptability. When production conditions change (such as replacing different plastic raw materials), the system can quickly re-optimize the parameter configuration and find a new balance point. At the same time, the differential privacy mechanism ensures that sensitive process parameter information will not be leaked during the parameter optimization process.

[0161] S3.3: Based on the target model parameters, train a temperature field distribution prediction model.

[0162] In S3.3, the system trains a temperature field distribution prediction model based on the optimized model parameters, i.e., the target model parameters. This model adopts a deep probabilistic graph network structure, including multiple random variable layers, which can capture both the spatial correlation and temporal dynamic characteristics of the temperature field. The specific structure can be as follows: the first layer processes short-term predictions, with a prediction range of 0 - 10 seconds, mainly relying on the local characteristics and immediate dynamics of the temperature field; the second layer processes medium-term predictions, with a prediction range of 10 - 60 seconds, considering both the spatial transfer characteristics of the temperature field and the influence of process parameters; the third layer processes long-term predictions, with a prediction range exceeding 60 seconds, focusing on the overall evolution trend of the temperature field. Model training adopts a sliding window strategy, that is, using the data of the last N time steps to predict the temperature distribution in the next M time steps. The system evaluates the model performance through cross-validation and dynamically adjusts the weights of the prediction time domain to ensure that the model can maintain good accuracy in different prediction ranges.

[0163] In addition, during the training process, the variational inference method is introduced to optimize the model parameters by minimizing the evidence lower bound (ELBO). Moreover, a dynamic learning rate adjustment mechanism is designed to adaptively adjust the learning step size according to the convergence of the loss function.

[0164] These steps form a complete process for constructing a prediction model. The reasonable construction of the prior distribution provides an initial knowledge basis for the model. Differential privacy and Bayesian optimization ensure the security and optimality of the model, while the hierarchical prediction structure enables the model to adapt to prediction requirements at different time scales. For example, in practical applications, short-term prediction can be used for real-time control adjustment, medium-term prediction can be used for process parameter optimization, and long-term prediction can be used for equipment maintenance planning. In addition, the system regularly updates the model parameters to adapt to changes in process conditions and the evolution of equipment performance. The prediction results of the model include not only the expected values of the temperature field but also the estimated uncertainty of the prediction, which provides a more comprehensive basis for subsequent control decisions.

[0165] S4: Perform feature fusion on the predicted die temperature results and the pressure data to construct the input features of the neural network controller, and train the neural network controller based on the input features and the positive definite curvature learning method, and generate the power control signals for the heaters in each area of the injection molding die and the flow control instructions for the water cooling system through the neural network controller;

[0166] In the embodiment of the present application, the construction of the neural network controller is first carried out. The controller adopts a dual-branch structure, with one branch processing the temperature prediction information and the other branch processing the pressure data. The inputs of the temperature branch include: the current temperature field distribution (16 measurement points), the predicted temperature change trend in the next 5 seconds, the temperature gradient information, and the hot spot distribution probability. The inputs of the pressure branch include: the real-time readings of each pressure sensor, the pressure change rate, and the pressure fluctuation characteristics. The two branches are connected through a feature fusion layer, and the fusion layer adopts an attention mechanism to automatically adjust the weights of the temperature and pressure features according to different working conditions. For example, at the initial stage of injection molding, the weight of the pressure feature is relatively high, while during the holding pressure stage, the weight of the temperature feature will increase accordingly.

[0167] The hidden layer of the network adopts a residual structure, which contains multiple residual blocks. Each residual block contains two fully connected layers and a skip connection, and this structure helps the network learn the subtle changes in the temperature field. The output layer is divided into two parts: the power control signals for the heaters (corresponding to the heaters in each area of the die) and the flow control instructions for the water cooling system (corresponding to different cooling circuits). The output layer uses the Sigmoid activation function to ensure that the control signals are within a reasonable range.

[0168] In terms of feature training, the system adopts a multi-stage training strategy. The first stage is feature preprocessing, which standardizes the temperature and pressure data and extracts time-series features. For example, for temperature data, in addition to the original value, the short-term (1 second), medium-term (5 seconds), and long-term (30 seconds) change rates are also calculated. The second stage is feature selection, which uses Lasso regression to identify the most influential feature combinations. The third stage is feature fusion, which learns the compact representation of temperature and pressure features through an autoencoder.

[0169] The application of the positive definite curvature learning method ensures the stability of the controller. First, the system constructs a loss function, including a control accuracy term, an energy consumption term, and a stability term. The control accuracy term measures the accuracy of temperature control, the energy consumption term optimizes the energy consumption of the heating and cooling systems, and the stability term ensures the smoothness of the control response through positive definite curvature constraints. During the training process, the system updates the network parameters through a second-order optimization method (such as a variant of Newton's method), while maintaining the positive definite curvature property of the loss function. This ensures that the controller can maintain stable response characteristics under various operating conditions.

[0170] The generation process of the control signal is divided into three steps: First, the controller calculates the ideal temperature field distribution based on the current state and the prediction result; then, it determines the control actions required to reach the target state through backpropagation; finally, it converts the control actions into specific execution instructions. Taking the heating control as an example, the system not only outputs the power value but also plans the timing pattern of heating, such as the pulse width modulation parameters. For the water cooling system, the control instructions include the flow rate and the cooling timing.

[0171] In practical applications, the controller shows good adaptability. For example, when it detects that the temperature in a certain area rises too fast, the system will reduce the power of the corresponding heater in advance and appropriately increase the flow rate of the adjacent water cooling circuit. Another example is that when switching different products in the mold, the controller can automatically adjust the control strategy according to the characteristics of the new product and quickly reach a stable production state.

[0172] The controller also has the ability of adaptive learning. During the production process, the system continuously collects data on the control effect and regularly updates the network parameters. This online learning mechanism enables the controller to adapt to the slow changes in the equipment performance, such as the reduction of the heater efficiency or the blockage of the water circuit. At the same time, the system will also record the control strategies in special cases to form an experience library for quick response to similar situations.

[0173] It should be noted especially the safety mechanism of the controller. The system sets multiple safety constraints, including temperature upper limit control, power gradual change limit, and emergency shutdown strategy. When detecting an abnormal situation, the controller will automatically switch to the safety mode to ensure the safety of the equipment first. These safety mechanisms are implemented through hard coding and are not affected by the neural network learning process.

[0174] As Figure 5 shown, step S4 may specifically include S4.1 to S4.5:

[0175] S4.1: By means of a sliding time window mechanism, construct the die temperature prediction results and real-time pressure data into a time series feature sequence;

[0176] Perform feature fusion on the temperature prediction results output from the foregoing steps and the real-time collected pressure data to construct the input feature space of the neural network controller. Specifically, the temperature prediction results contain the temperature mean and uncertainty estimation of each region at future moments, and the pressure data includes the real-time pressure values and pressure change rates at key positions of the die. The system adopts a sliding time window mechanism to combine historical data and prediction data into a time series feature sequence, and at the same time introduces an attention mechanism to weight the features to highlight the data features at important moments.

[0177] Specifically, a main time window of 60 seconds can be adopted, with a sliding step of 1 second, and multiple sub-windows are set within the main window to capture features at different time scales. For the temperature prediction results, each prediction point contains the temperature change trend in the next 5 seconds, and the sampling frequency is 10Hz, so a sequence of 50 prediction points is formed. For the pressure data, a sampling frequency of 10Hz is also adopted to construct a pressure sequence containing 600 sampling points. For example, the features of a certain measurement point at time t may include: the current temperature of 165°C, the predicted temperature sequence [165.2, 165.5, 165.8, 166.0, 166.1]°C in the next 5 seconds, and the corresponding pressure sequence [45.2, 45.5, 45.8, 46.0, 46.1] MPa. The system will also calculate the statistical features of these sequences, such as the mean, standard deviation, and change rate.

[0178] S4.2: Through the attention mechanism, weight the time series feature sequence to obtain a weighted feature sequence;

[0179] In S4.2, the system implements a multi-level attention mechanism. First is the attention in the time dimension, and the system assigns weights to the features at different time points according to the prediction accuracy. For example, if it is found that the temperature prediction at a certain time point is particularly accurate, the weight of this point will be increased accordingly. Second is the attention in the spatial dimension, and the system assigns weights according to the importance of different regions. For example, hot spots and regions with large temperature gradients will obtain higher weights. Finally is the attention in the feature dimension, and the system will dynamically adjust the weight ratio of temperature features and pressure features. In the initial stage of injection molding, a weight of 0.7 may be given to the pressure feature and a weight of 0.3 to the temperature feature; while in the holding pressure stage, it may be adjusted to a weight of 0.4 for the pressure feature and a weight of 0.6 for the temperature feature.

[0180] S4.3: Perform eigenvector transformation on the weighted feature sequence to generate the input features of the neural network controller.

[0181] In S4.3, the system converts the weighted sequence into the input features of the controller through a feature transformation network. The transformation network consists of three key modules: a temporal encoding module that converts the time series into a vector of fixed dimension; a spatial encoding module that captures the spatial relationships between different measurement points; and a feature fusion module that combines the temporal features and spatial features into the final input vector. For example, for a mold with 16 measurement points, a 256-dimensional feature vector may be finally generated, where the first 128 dimensions encode temperature-related information and the last 128 dimensions encode pressure-related information.

[0182] S4.4: Construct a loss function based on the Riemann metric according to the input features, where the loss function includes a temperature control error term, a pressure balance term, and an energy consumption constraint term, and obtain a curvature regularization term to get an optimized objective function;

[0183] In S4.4, the system constructs a composite loss function based on the Riemann metric. The temperature control error term uses the weighted Euclidean distance to measure the deviation between the actual temperature and the target temperature, and the weights are dynamically adjusted according to the importance of different regions. The pressure balance term uses the condition number of the covariance matrix as an index to ensure uniform pressure distribution in each region. The energy consumption constraint term includes the sum of the squares of the heating power and the cooling power, as well as their rates of change. For example, if the heating power in a certain region changes frequently and significantly, it will be penalized by the energy consumption constraint term. The curvature regularization term is constructed by calculating the Riemann curvature in the parameter space to ensure the smoothness of the control response.

[0184] S4.5: Calculate the Riemann gradient in the tangent space of the parameters according to the optimized objective function, and project the updated parameters back to the manifold space satisfying the positive definite constraint through the exponential map to obtain the optimized network parameters.

[0185] In this embodiment, considering the particularity of the positive definite curvature constraint, the traditional gradient descent method may lead to a violation of the positive definiteness constraint after parameter update. To solve this problem, a parameter update strategy based on Riemannian optimization is adopted in this embodiment. Specifically, during each parameter update, the Riemannian gradient is first calculated in the tangent space of the parameter. This gradient takes into account the geometric structure of the parameter space. Then, the updated parameter is projected back to the manifold space that satisfies the positive definite constraint through the exponential map. For example, if a certain weight matrix after update does not satisfy positive definiteness, the system will project it to the nearest positive definite matrix. This process ensures that the controller can maintain stability under various working conditions. In addition, to improve the training efficiency, the system also introduces a momentum term and an adaptive learning rate mechanism, which can dynamically adjust the update step size according to the gradient change. Specifically, for a typical control cycle: first calculate the Riemannian gradient based on the current state, then use a retry mechanism to ensure finding an appropriate learning rate, and finally update the parameter through the exponential map.

[0186] An important feature of this training method is its stability. By optimizing on the Riemannian manifold, the system can avoid the numerical instability problems that may be encountered in traditional gradient descent. At the same time, the positive definite constraint ensures that the response of the controller always evolves in the direction of reducing the error. For example, if a certain control action causes the temperature deviation to increase, the system will immediately adjust the parameters to make the subsequent control actions more conservative. In addition, the system will regularly verify the stability of the controller. If potential instability factors are found, it will automatically adjust the regularization strength.

[0187] This will be illustrated by a specific injection mold temperature control scenario:

[0188] Suppose a mold with a 4×4 grid layout, having a total of 16 temperature monitoring points and 8 pressure monitoring points. In S4.4, the loss function constructed by the system based on the Riemannian metric contains three main terms. The temperature control error term reflects the deviation between the actual temperature and the target temperature, and adopts a weighted form:

[0189]

[0190] where wi is the weight of each measurement point (hotter regions have higher weights), Ti is the measured temperature, and Ti_target is the target temperature. For example, when the temperature in the upper right corner region is detected to be 175 °C, the target temperature is 170 °C, and the weight of this region is 1.5, the contribution of this part of the region to the total error is 1.5 (175 - 170)² = 37.5.

[0191] The pressure balance term and the energy consumption constraint term form a composite constraint. The pressure balance is evaluated by the standard deviation of each pressure measurement point, and the energy consumption constraint considers the sum of the squares of the heating power and the cooling power. For example, in a certain control, the system records that the powers of 4 heating zones are [800W, 750W, 720W, 780W] respectively, and the flow rates of 2 cooling circuits are [2.5L / min, 2.3L / min] respectively. The system will calculate the variances and change rates of these control variables, and larger fluctuations will be imposed with larger penalty terms.

[0192] In S4.5, the parameter optimization process is divided into three stages. The first stage is initialization, where the system randomly generates initial parameters that satisfy the positive definite condition in a 16×16 weight matrix space. For example, the initial value of the weight matrix of a certain layer may be the identity matrix plus a small random perturbation. The second stage is gradient calculation and parameter update, where the system calculates the Riemannian gradient in the tangent space of the parameters. Suppose in a certain iteration, the system detects that the temperature in the upper right corner area continues to be high, and the corresponding weight update will enhance the cooling control weights related to this area while reducing the heating control weights.

[0193] The third stage is parameter projection to ensure that the updated parameters still satisfy the positive definite constraint. For example, the eigenvalues of a 4×4 weight sub - matrix obtained after a certain update are [1.2, 0.8, -0.1, 0.5], where a negative eigenvalue -0.1 appears. The system will project it to the nearest positive definite matrix through the exponential mapping to obtain a new eigenvalue distribution [1.2, 0.8, 0.1, 0.5]. This projection ensures the stability of the controller.

[0194] In practical applications, this optimization method exhibits good convergence and robustness. For example, during the mold heating stage, the controller can smoothly increase the temperature to the target value and avoid overshoot. Specifically, when the mold is heated from room temperature to 165℃, the heating curve is approximately exponential, automatically reducing the heating power when approaching the target temperature, and finally controlling the temperature deviation within ±0.5℃. At the same time, the temperature difference between different regions is controlled within 2℃, and the pressure fluctuation does not exceed 5%.

[0195] This optimization method based on Riemannian geometry is particularly suitable for dealing with control problems with strong coupling characteristics such as temperature fields. For example, when a temperature anomaly is detected in a certain area, the controller will not only adjust the control parameters of this area but also synchronously optimize the parameters of adjacent areas to achieve an overall coordinated control effect. During the production process, if the efficiency of a certain heater decreases, the system will automatically adjust the relevant weights to compensate for this performance decay and maintain the overall control effect.

[0196] The system also has an adaptive feature. When process conditions change (such as when different materials are replaced), the controller can quickly adjust parameters while maintaining stability. For example, when switching from PET material to ABS material, the system will automatically adapt to the thermal characteristics of the new material within 2 - 3 production cycles, adjust the control strategy, and ensure stable product quality. This adaptive process always remains within the positive definite constraint range, avoiding the risk of controller instability.

[0197] S5: Based on the power control signal of the heater and the flow control instruction of the water cooling system, dynamically control the temperature field of the injection molding die through an adjustable gap compensation mechanism;

[0198] Step S5 may specifically include steps S5.1 to S5.3:

[0199] S5.1: Based on the power control signal of the heater, modulate the power of the heaters in each area of the injection molding die through a thyristor module, and based on the flow control instruction of the water cooling system, control the cooling water flow of the water cooling system in each area of the injection molding die through a high-precision proportional valve;

[0200] In this embodiment, based on the power control signal of the heaters in each area and the flow control instruction of the water cooling system, the temperature field of the injection molding die is dynamically controlled. Specifically, an industrial-grade PLC controller is used as the execution unit and communicates with each actuator through a high-speed fieldbus. For the heating system, an SCR (thyristor) module is used for precise power modulation, supporting an adjustment accuracy of 0.1%. For example, when the controller outputs that a certain area requires 75% power, the thyristor module will precisely control the conduction angle within each power frequency cycle to achieve accurate power modulation. In addition, to avoid electromagnetic interference, the system is triggered near the zero crossing point and a minimum conduction time limit (such as 1 ms) is set. For the water cooling system, a high-precision proportional valve is used to control the cooling water flow, and the control resolution reaches 0.1 L / min. Specifically, according to the flow control instruction, the proportional valve is driven by a PWM signal to achieve precise flow adjustment. For example, when the cooling water flow in a certain area needs to be reduced from 2.5 L / min to 1.8 L / min, the proportional valve will smoothly complete the adjustment within 100 ms, avoiding temperature fluctuations caused by sudden flow changes. It should be noted that the system also designs a hardware redundancy mechanism to ensure that the basic temperature control function can still be maintained when a single execution unit fails.

[0201] S5.2: Real-time monitor the displacement change of the adjustable gap compensation mechanism, and adjust the gap size of the cavity of the injection molding die based on the displacement change;

[0202] In this embodiment, the intelligent servo driver is used to monitor and adjust the displacement change of the gap compensation mechanism in real time. The gap compensation mechanism adopts a high-precision electric push rod design and has a positioning accuracy at the micron level. During actual operation, the system first calculates the theoretical compensation amount based on the thermal deformation model, and then realizes precise positioning through closed-loop displacement feedback. Specifically, the displacement of the gap compensation mechanism is monitored in real time by a high-precision displacement sensor (resolution 0.001 mm), the displacement data sampling frequency is 100 Hz, and high-frequency noise is removed through low-pass filtering. When it is detected that the change in the cavity gap exceeds a preset threshold (such as 0.02 mm), the system will initiate compensation adjustment. For example, during the injection molding process, if it is detected that the gap in a certain area has increased by 0.03 mm due to thermal expansion, the compensation mechanism will perform precise compensation within 200 ms to adjust the gap back to the target value. This process is gradual to avoid sudden mechanical movement from affecting the molding quality. At the same time, the system will record the gap change rules under different working conditions for predictive compensation.

[0203] In addition, in order to prevent mold damage caused by over-compensation, multiple safety protection mechanisms are set up, including functions such as displacement limit, torque protection, and collision detection. It is particularly worth mentioning that the system has also established a mapping relationship database between thermal deformation and gap compensation, which can quickly determine the optimal compensation strategy based on historical experience.

[0204] S5.3: Based on the temperature data and pressure data of the injection molding die collected in real time, the PID parameters are adjusted through the fuzzy adaptive algorithm, and the prediction period and control time domain of the model predictive controller are optimized based on the reinforcement learning method to obtain the optimized control parameters, and when it is detected that the process parameters meet the preset change conditions, the parameter re-optimization process is triggered.

[0205] In this embodiment, a complete set of dynamic parameter adjustment mechanisms is established. Through the temperature and pressure data collected in real time, the system can evaluate the current control effect and dynamically optimize the control parameters accordingly. Specifically, the fuzzy adaptive algorithm is used to adjust the PID parameters online, and at the same time, the prediction period and control time domain of the model predictive controller are optimized based on the reinforcement learning method. When it is detected that the process parameters change significantly, the system will automatically trigger the parameter re-optimization process to ensure that the control performance always remains in the best state. In addition, the system also has the function of self-learning process parameters and can continuously improve the control strategy library according to historical production data.

[0206] Specifically, in this embodiment, double-layer adaptive control optimization is achieved. First, it is the fuzzy adaptive adjustment of PID parameters. The system divides the working state into 9 fuzzy regions (such as "large positive error - fast positive change rate", "small error - slow negative change rate", etc.) according to the temperature error and the change rate of the error. Within each region, the system adjusts the PID parameters according to the preset fuzzy rules. For example, when it is detected that the temperature error is large and decreasing rapidly, the system will appropriately reduce the proportional gain and increase the integral time to avoid overshoot. These adjustments are smooth to ensure the stability of the control process.

[0207] Second, it is the model predictive control optimization based on reinforcement learning. The system takes the control effect as the reward signal, including three aspects: temperature control accuracy, energy consumption efficiency, and response speed. Through the Q-learning algorithm, the system learns the optimal prediction period and control time domain under different working conditions. For example, in the rapid heating stage, the system may choose a shorter prediction period (such as 2 seconds) and control time domain (such as 1 second) to obtain a fast response; while in the stable production stage, it may adopt a longer prediction period (such as 10 seconds) and control time domain (such as 5 seconds) to improve stability and energy-saving effect.

[0208] The system sets multiple trigger conditions to start the parameter re-optimization process, including: process parameter changes exceeding the set threshold (such as the temperature set value changing by more than 5 °C), control performance degradation (such as the temperature fluctuation increasing by 20%), equipment state change (such as changing the mold or material), etc. When the re-optimization is triggered, the system will perform parameter optimization calculations in the background while maintaining the current control. For example, when it is detected that the efficiency of a certain heater drops by 15%, the system will re-optimize the control parameters in this area, including PID parameters and predictive control parameters, and appropriately adjust the parameters in adjacent areas to maintain the overall control effect.

[0209] This adaptive optimization mechanism enables the system to cope with various working condition changes. For example, when replacing products with different thicknesses during the production process, the system can automatically adjust to the optimal control state within 2 - 3 production cycles. Another example is that when the thermal characteristics of the mold change after long-term use, the system can maintain the control performance through continuous parameter optimization. At the same time, the system will record the optimization experience under different working conditions and establish an experience database for quick response to similar situations.

[0210] In practical applications, this multi-level control optimization strategy exhibits excellent adaptability and stability. For example, in a certain production, when the moisture content of the material suddenly increases, resulting in an increase in the forming pressure fluctuation, the system can maintain the stability of the temperature field while adjusting the control parameters to suppress the pressure fluctuation and ensure the stable quality of the product. The system also has a predictive maintenance function. By analyzing the long-term change trend of the control parameters, potential equipment problems can be detected in advance, such as a decrease in the heater efficiency or a blockage in the water circuit.

[0211] As shown Figure 6 in the figure, the embodiment of the present application further provides an intelligent temperature regulation system 500 for an injection molding die, specifically including:

[0212] A data acquisition module 501, configured to acquire temperature data and pressure data of each area of the injection molding die, and generate a real-time multi-source sensor data stream;

[0213] A feature recognition module 502, configured to extract features from the real-time multi-source sensor data stream through a Transformer model and an EM algorithm to obtain temperature field distribution features;

[0214] A temperature prediction module 503, configured to train a temperature field distribution prediction model based on the temperature field distribution features and a differential private Bayesian optimization algorithm, and output a die temperature prediction result through the temperature field distribution prediction model;

[0215] A control strategy module 504, configured to perform feature fusion on the die temperature prediction result and the pressure data, construct input features of a neural network controller, train the neural network controller based on the input features and a positive definite curvature learning method, and generate a power control signal for heaters of each area of the injection molding die and a flow control instruction for a water cooling system through the neural network controller;

[0216] An execution control module 505, configured to perform dynamic control on the temperature field of the injection molding die through an adjustable gap compensation mechanism based on the power control signal of the heater and the flow control instruction of the water cooling system.

[0217] The system of the embodiment of the present disclosure can execute the method provided by the embodiment of the present disclosure, and its implementation principle is similar. The actions performed by each module in the device of each embodiment of the present disclosure correspond to the steps in the method of each embodiment of the present disclosure. For the detailed function descriptions of each module of the device, reference can specifically be made to the descriptions in the corresponding method shown above, and details are not described herein again.

[0218] Finally, it should be noted that the above-described embodiments are only specific implementation manners of the present disclosure, used to illustrate the technical solutions of the present disclosure, rather than limiting it. The protection scope of the present disclosure is not limited thereto. Although the present disclosure has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that any person skilled in the art within the technical scope disclosed by the present disclosure can still modify the technical solutions described in the foregoing embodiments, or can easily think of changes, or perform equivalent replacements on some of the technical features; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present disclosure, and should all be covered within the protection scope of the present disclosure. Therefore, the protection scope of the present disclosure should be subject to the protection scope of the claims.

Claims

1. An intelligent temperature adjustment method for an injection molding mold, characterized in that: include: Collect temperature and pressure data from all areas of the injection molding mold to generate real-time multi-source sensor data streams; By using the Transformer model and the EM algorithm, feature extraction is performed on the real-time multi-source sensor data stream to obtain temperature field distribution features; Based on the temperature field distribution characteristics and the differential private Bayesian optimization algorithm, a temperature field distribution prediction model is trained and obtained, and a mold temperature prediction result is output through the temperature field distribution prediction model; Performing feature fusion on the mold temperature prediction result and the pressure data to construct input features of a neural network controller, and training the neural network controller based on the input features and a positive curvature learning method to obtain the neural network controller, and generating power control signals for heaters in each area of ​​the injection molding mold and flow control instructions for a water cooling system through the neural network controller; Based on the power control signal of the heater and the flow control instruction of the water cooling system, the temperature field of the injection molding mold is dynamically controlled through an adjustable gap compensation mechanism; Wherein, the training to obtain the temperature field distribution prediction model based on the temperature field distribution characteristics and the differential private Bayesian optimization algorithm includes: Based on the temperature field distribution characteristics, constructing a priori distribution of a Bayesian probability model; The model parameters are randomly perturbed through the differential privacy mechanism, and the target model parameters are searched through the Bayesian optimization method; Based on the target model parameters, a temperature field distribution prediction model is obtained by training; The prior distribution of the Bayesian probability model is constructed based on the temperature field distribution characteristics, including: By multi-dimensional Gaussian distribution modeling and historical data analysis, an initial prior distribution is obtained based on the temperature gradient vector field and the hot spot distribution characteristic map in the temperature field distribution characteristics; Based on the initial prior distribution, the parameter distribution of different working conditions is automatically learned through a hierarchical Bayesian structure to generate a hierarchical prior model; Constructing a Bayesian probability model based on the hierarchical prior model and outputting prior distribution parameters; Furthermore, the method of randomly perturbing the model parameters by using the differential privacy mechanism and searching the target model parameters by using the Bayesian optimization method includes: Add random noise to the prior distribution parameters through the Laplace mechanism, and dynamically adjust the privacy budget based on the parameter sensitivity to obtain the disturbed parameters; A proxy model is constructed by Gaussian process regression, parameter performance evaluation is performed on the perturbed parameters, and a candidate parameter set is obtained by exploring and utilizing a balance strategy, wherein the evaluation indicators of the proxy model include prediction accuracy and computational efficiency; For the candidate parameter set, the model parameters are optimized by using a variational inference method and a method for minimizing the lower bound of evidence, and the target model parameters are output; The method of obtaining the neural network controller based on the input feature and the positive curvature learning method training includes: According to the input features, a loss function based on the Riemann metric is constructed, wherein the loss function includes a temperature control error term, a pressure balance term, and an energy consumption constraint term, and a curvature regularization term is obtained to obtain an optimization objective function; According to the optimization objective function, the Riemann gradient is calculated in the tangent space of the parameters, and the updated parameters are projected back to the manifold space satisfying the positive definite constraint through exponential mapping to obtain the optimized network parameters.

2. The method according to claim 1, characterized in that The method of collecting temperature data and pressure data of each area of ​​the injection molding mold to generate a real-time multi-source sensor data stream includes: The temperature data collected by the temperature sensor and the pressure data collected by the pressure sensor are converted into digital form through a high-speed data acquisition card to obtain digital signals; Performing signal filtering, data calibration, and outlier elimination on the digital signal to generate preprocessed data; A distributed data caching mechanism is established to synchronize and align the pre-processed data according to timestamps to generate a real-time multi-source sensor data stream.

3. The method according to claim 1, characterized in that The method of extracting features from the real-time multi-source sensor data stream by using the Transformer model and the EM algorithm to obtain temperature field distribution features includes: By using a position encoder of a Transformer model, the real-time multi-source sensor data stream is added with time series position information to generate an initial feature representation; Based on the initial feature representation, a multi-head attention mechanism is used to calculate the similarity between the query matrix, the key matrix and the value matrix, and generate an attention weight matrix for reflecting the spatiotemporal correlation of the data; Using a residual connection structure, the attention weight matrix is ​​weightedly fused with the initial feature representation to obtain an enhanced feature representation; The enhanced feature representation is clustered and analyzed by an EM algorithm to obtain the temperature field distribution feature.

4. The method according to claim 3, characterized in that The step of performing cluster analysis on the enhanced feature representation by using an EM algorithm to obtain the temperature field distribution feature includes: Based on the enhanced feature representation, K cluster centers are randomly initialized to obtain initial clustering parameters, wherein the K value is adaptively determined by the silhouette coefficient; According to the initial clustering parameters, alternately performing the steps of calculating the probability of belonging of data points and updating the cluster centers, and dynamically adjusting the parameter updating step size through an adaptive learning rate mechanism until a target clustering result is obtained; The temperature field distribution characteristics are obtained through the target clustering results, wherein the temperature field distribution characteristics include a temperature gradient vector field, a hot spot distribution characteristic map, and a temperature fluctuation spectrum.

5. The method according to claim 1, characterized in that The step of fusing the mold temperature prediction result and the pressure data to construct input features of a neural network controller includes: The mold temperature prediction result and the real-time pressure data are constructed into a time series feature sequence through a sliding time window mechanism; By using an attention mechanism, the temporal feature sequence is weighted to obtain a weighted feature sequence; The weighted feature sequence is converted into a feature vector to generate input features of a neural network controller.

6. The method according to claim 1, characterized in that The method dynamically controls the temperature field of the injection molding die through an adjustable gap compensation mechanism based on the power control signal of the heater and the flow control instruction of the water cooling system, including: Based on the power control signal of the heater, the power of the heater in each area of ​​the injection molding mold is modulated through a thyristor module, and based on the flow control instruction of the water cooling system, the cooling water flow of the water cooling system in each area of ​​the injection molding mold is controlled through a high-precision proportional valve; Real-time monitoring of the displacement change of the adjustable gap compensation mechanism, and adjusting the gap size of the cavity of the injection molding mold based on the displacement change; Based on the temperature data and pressure data of the injection molding mold collected in real time, PID parameters are adjusted through a fuzzy adaptive algorithm, and the prediction period and control time domain of the model predictive controller are optimized based on the reinforcement learning method to obtain the optimized control parameters. When it is detected that the process parameters meet the preset change conditions, the parameter re-optimization process is triggered.

7. An intelligent temperature control system for injection molding molds, characterized in that: include: Data acquisition module, used to collect temperature and pressure data of various areas of the injection molding mold and generate real-time multi-source sensor data stream; A feature recognition module is used to extract features from the real-time multi-source sensor data stream through a Transformer model and an EM algorithm to obtain temperature field distribution features; A temperature prediction module, used for training a temperature field distribution prediction model based on the temperature field distribution characteristics and a differential private Bayesian optimization algorithm, and outputting a mold temperature prediction result through the temperature field distribution prediction model; A control strategy module, used to perform feature fusion on the mold temperature prediction result and the pressure data, construct input features of a neural network controller, and obtain the neural network controller through training based on the input features and a positive curvature learning method, and generate power control signals for heaters in each area of ​​the injection molding mold and flow control instructions for a water cooling system through the neural network controller; An execution control module, configured to dynamically control the temperature field of the injection molding die through an adjustable gap compensation mechanism based on a power control signal of the heater and a flow control instruction of the water cooling system; Wherein, the training to obtain the temperature field distribution prediction model based on the temperature field distribution characteristics and the differential private Bayesian optimization algorithm includes: Based on the temperature field distribution characteristics, constructing a priori distribution of a Bayesian probability model; The model parameters are randomly perturbed through the differential privacy mechanism, and the target model parameters are searched through the Bayesian optimization method; Based on the target model parameters, a temperature field distribution prediction model is obtained by training; The prior distribution of the Bayesian probability model is constructed based on the temperature field distribution characteristics, including: By multi-dimensional Gaussian distribution modeling and historical data analysis, an initial prior distribution is obtained based on the temperature gradient vector field and the hot spot distribution characteristic map in the temperature field distribution characteristics; Based on the initial prior distribution, the parameter distribution of different working conditions is automatically learned through a hierarchical Bayesian structure to generate a hierarchical prior model; Constructing a Bayesian probability model based on the hierarchical prior model and outputting prior distribution parameters; Furthermore, the method of randomly perturbing the model parameters by using the differential privacy mechanism and searching the target model parameters by using the Bayesian optimization method includes: Add random noise to the prior distribution parameters through the Laplace mechanism, and dynamically adjust the privacy budget based on the parameter sensitivity to obtain the disturbed parameters; A proxy model is constructed by Gaussian process regression, parameter performance evaluation is performed on the perturbed parameters, and a candidate parameter set is obtained by exploring and utilizing a balance strategy, wherein the evaluation indicators of the proxy model include prediction accuracy and computational efficiency; For the candidate parameter set, the model parameters are optimized by using a variational inference method and a method for minimizing the lower bound of evidence, and the target model parameters are output; The method of obtaining the neural network controller based on the input feature and the positive curvature learning method training includes: According to the input features, a loss function based on the Riemann metric is constructed, wherein the loss function includes a temperature control error term, a pressure balance term, and an energy consumption constraint term, and a curvature regularization term is obtained to obtain an optimization objective function; According to the optimization objective function, the Riemann gradient is calculated in the tangent space of the parameters, and the updated parameters are projected back to the manifold space satisfying the positive definite constraint through exponential mapping to obtain the optimized network parameters.

Images