A mold temperature synchronous control method and system

Abstract

The application discloses a kind of mould temperature synchronous control method and system, comprising: first standardization accesses the detection time series data stream of the detection point of the mould to be regulated and obtains process condition parameter.Response to reach temperature synchronous control condition, according to the real-time thermal state feature of time point M to L, reference state feature in state cache unit is updated, and dynamic thermal state feature is obtained.Reference state feature is generated by history data modeling before time point M.The estimated temperature control target value is generated based on thermal state feature, process condition parameter, first control strategy parameter and pre-training temperature synchronous control model.The control model integrates heat sensing feature extraction network, feature mapping network and control model: heat sensing feature extraction network extracts thermal state feature from time series data;Feature mapping network aligns feature scale;Control model outputs target value by fusing mapped feature, process parameter, previous control result and strategy parameter, to realize accurate synchronous control of mould multi-point temperature.

Description

Technical Field

[0001] This invention relates to the field of industrial control technology, and more specifically, to a method and system for synchronous control of mold temperature. Background Technology

[0002] Mold temperature control directly impacts product quality consistency in manufacturing processes such as injection molding and die casting. Traditional temperature control methods rely on fixed thresholds or static models, which struggle to adapt to the dynamic thermal behavior differences across multiple points in the mold. Especially in large and complex mold scenarios, uneven temperature leads to frequent problems such as shrinkage deformation and surface defects. Existing technologies need to combine historical data with real-time process parameters to construct dynamic thermal models, achieving simultaneous temperature control across multiple regions to ensure manufacturing accuracy. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for synchronous control of mold temperature.

[0004] In a first aspect, embodiments of the present invention provide a method for synchronous control of mold temperature, comprising:

[0005] The detection points of the mold to be adjusted and the process condition parameters of the detection points of the mold to be adjusted are obtained, and the detection points of the mold to be adjusted are accessed in a standardized manner using the detection timing data stream;

[0006] In response to reaching the temperature synchronization control condition, the thermal state characteristics of the detection point of the mold to be controlled are acquired. The thermal state characteristics of the detection point of the mold to be controlled are obtained by updating the state distribution model of at least one reference state characteristic stored in the state cache unit based on the thermal state characteristics of the detection point of the mold to be controlled from time point M to time point L. The at least one reference state characteristic is obtained by updating the state distribution model of the thermal state characteristics of the detection point of the mold to be controlled before time point M. The time point L is the monitoring period of the detection point of the mold to be controlled when the detection point of the mold to be controlled is identified. The L is less than or equal to the monitoring period of the detection point of the mold to be controlled.

[0007] Based on the thermal state characteristics of the detection point of the mold to be regulated, the process condition parameters of the detection point of the mold to be regulated, the first control strategy parameters, and the pre-trained temperature synchronization regulation model, the estimated temperature regulation target value of the detection point of the mold to be regulated is determined. The temperature synchronization regulation model includes a thermal feature extraction network, the state buffer unit, a feature mapping network, and a regulation model.

[0008] The thermal feature extraction network is used to extract the thermal state features corresponding to the detection time series data. The feature mapping network is used to perform feature space mapping and feature scale transformation on the thermal state features, aligning the thermal state features to the input scale of the control model to obtain the thermal state features after feature space mapping. The control model is used to generate the estimated temperature control target value of the detection point of the mold to be controlled by taking the thermal state features after feature space mapping, the process condition parameters of the detection point of the mold to be controlled, the result of the previous round of temperature synchronization control of the detection point of the mold to be controlled, and the first control strategy parameters as input.

[0009] In a second aspect, embodiments of the present invention provide a server system, including a server, the server being used to execute the method described in the first aspect.

[0010] Compared to existing technologies, the beneficial effects provided by this invention include: The mold temperature synchronization control method and system disclosed in this invention pertains to the field of industrial control technology. It obtains process condition parameters by standardizing the access to the detection time-series data stream of the mold's detection points. Upon reaching the temperature synchronization control condition, the reference state features in the state cache unit are updated based on the real-time thermal state characteristics from time points M to L, obtaining dynamic thermal state characteristics. The reference state features are generated by modeling historical data before time point M. Based on the thermal state characteristics, process condition parameters, first control strategy parameters, and a pre-trained temperature synchronization control model, an estimated temperature control target value is generated. This control model integrates a thermal feature extraction network, a feature mapping network, and the control model: the thermal feature extraction network extracts thermal state features from the time-series data; the feature mapping network aligns the feature scale; and the control model fuses the mapped features, process parameters, previous control results, and strategy parameters to output the target value, achieving precise synchronous temperature control at multiple points on the mold. Attached Figure Description

[0011] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as limiting the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a flowchart illustrating the steps of the mold temperature synchronization control method provided in an embodiment of the present invention.

[0013] Figure 2 A schematic block diagram of the structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0015] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0016] In order to solve the technical problems mentioned in the background art Figure 1 This is a flowchart illustrating the mold temperature synchronization control method provided in this embodiment. The mold temperature synchronization control method will be described in detail below.

[0017] Step S201: Obtain the detection points of the mold to be adjusted and the process condition parameters of the detection points of the mold to be adjusted. The detection points of the mold to be adjusted are accessed in a standardized manner using the detection timing data stream.

[0018] Step S202: In response to reaching the temperature synchronization control condition, the thermal state characteristics of the detection point of the mold to be controlled are acquired. The thermal state characteristics of the detection point of the mold to be controlled are obtained by updating the state distribution model of at least one reference state characteristic stored in the state cache unit based on the thermal state characteristics of the detection point of the mold to be controlled from time point M to time point L. The at least one reference state characteristic is obtained by updating the state distribution model of the thermal state characteristics of the detection point of the mold to be controlled before time point M. The time point L is the monitoring period of the detection point of the mold to be controlled when the detection point of the mold to be controlled is identified. The L is less than or equal to the monitoring period of the detection point of the mold to be controlled.

[0019] Step S203: Based on the thermal state characteristics of the detection point of the mold to be regulated, the process condition parameters of the detection point of the mold to be regulated, the first control strategy parameters, and the pre-trained temperature synchronization regulation model, determine the estimated temperature regulation target value of the detection point of the mold to be regulated. The temperature synchronization regulation model includes a thermal feature extraction network, the state buffer unit, a feature mapping network, and a regulation model.

[0020] The thermal feature extraction network is used to extract the thermal state features corresponding to the detection time series data. The feature mapping network is used to perform feature space mapping and feature scale transformation on the thermal state features, aligning the thermal state features to the input scale of the control model to obtain the thermal state features after feature space mapping. The control model is used to generate the estimated temperature control target value of the detection point of the mold to be controlled by taking the thermal state features after feature space mapping, the process condition parameters of the detection point of the mold to be controlled, the result of the previous round of temperature synchronization control of the detection point of the mold to be controlled, and the first control strategy parameters as input.

[0021] In this embodiment of the invention, exemplarily, the following describes in detail the mold temperature synchronization control method of the present invention, using a server as the execution entity, in conjunction with the actual production scenario of an automotive bumper injection mold. The mold includes four cavity detection points (denoted as P1 to P4), each equipped with a contact temperature sensor (sampling frequency 1Hz), a pressure sensor (sampling frequency 0.5Hz), and a thermocouple. The server communicates in real-time with the field PLC system via industrial Ethernet (Profinet protocol) to achieve detection data acquisition and control command issuance. The server hardware configuration includes an Intel Xeon Gold 6248 processor, 64GB of memory, and a 1TB SSD. The software environment includes Python 3.8 (with the PyTorch deep learning framework), a Redis database (as a state cache unit), and an OPCUA data gateway for standardizing the time-series data streams accessing each detection point.

[0022] Obtain the inspection points and process condition parameters of the mold to be adjusted:

[0023] The server first accesses the real-time timing data of each detection point through the OPCUA data gateway. This data is transmitted to the server's local data processing module in a standardized format (including fields such as point identifier, timestamp, temperature value, and pressure value). For example, for the P1 cavity detection point, the server receives the following timing data sample: point identifier "P1", timestamp "2023-10-15 08:00:00.000", temperature value 176.2℃, pressure value 85.3MPa, and injection cycle count 1250. The server preprocesses this raw data, including outlier removal (such as filtering invalid data with temperatures >300℃ or <-10℃), timing alignment (filling in missing sampling points based on timestamps), and normalization (mapping physical quantities such as temperature and pressure to the [0,1] interval) to ensure that the data meets the requirements of subsequent model input.

[0024] Simultaneously, the server retrieves the mold's process condition parameters from the factory's MES system, specifically including three categories: basic mold parameters, injection molding process parameters, and target temperature parameters. The basic mold parameters are: material is H13 hot work die steel (converted to a feature vector [1,0,0] through unique thermal encoding), and cavity dimensions are 200mm long, 150mm wide, and 40mm deep (normalized to [0.9,0.7,0.5]). The injection molding process parameters are: material is ABS resin (melt index 25g / 10min, encoded as [0,1,0]), injection cycle 45s (normalized to 0.6), and holding pressure 80bar (normalized to 0.7). The target temperature parameters are: target cavity surface temperature 180±2℃, and hot runner temperature 210±3℃. After vectorizing these process condition parameters, the server merges them into a 32-dimensional feature vector (10-dimensional mold parameters + 22-dimensional process parameters), storing it in a local MySQL database for subsequent use by the temperature synchronization control model.

[0025] Optionally, determining whether the temperature synchronization control conditions have been met and acquiring thermal state characteristics includes:

[0026] The server monitors the temperature fluctuations at each detection point in real time. Preset temperature synchronization control trigger conditions include: real-time temperature deviating from the target temperature threshold by ±5℃, three consecutive temperature fluctuations exceeding 2℃ within 10 seconds, or a process switching event (such as changing injection molding materials or adjusting the injection cycle). In this scenario, at 08:15:00 on October 15, 2023 (time point L, i.e., the end time of the current monitoring period), the server detected that the real-time temperature at point P1 remained at 172℃ for one minute, deviating from the target temperature of 180℃ by -8℃, thus meeting the trigger condition of "real-time temperature deviating from the target temperature threshold by ±5℃". The server then initiated the temperature synchronization control process and began acquiring the thermal characteristics of point P1.

[0027] The hot state characteristics of point P1 are obtained by dynamically updating the reference state characteristics stored in the state cache unit. The specific process is as follows: First, the server needs to construct the initial reference state characteristics based on the historical data before time point M (the historical data cutoff point before the trigger control, set as 2023-10-15 07:00:00, i.e. 1 hour before the trigger control). Then, the real-time data from time point M to time point L is used to model and update the state distribution of the reference state characteristics.

[0028] For constructing the initial reference state features, the server retrieves historical time-series data (7 × 24 × 3600 = 604,800 sampling points) from the local database for point P1 7 days prior to time point M. This data includes parameters such as temperature, pressure, and injection cycle counts. The server inputs this historical data into the thermal feature extraction network in the temperature synchronization control model (this network is an LSTM network, containing a 128-dimensional input layer, a 64-dimensional hidden layer, and a 32-dimensional output layer, with tanh as the activation function), extracting a 32-dimensional thermal state feature sequence. Subsequently, the server uses a Gaussian mixture model (GMM) to model the state distribution of this thermal state feature sequence, setting three mixture components (corresponding to the mold's "low-temperature stable state," "heating transition state," and "high-temperature stable state," respectively). The EM algorithm is used to estimate the weights, mean vectors, and covariance matrices of each mixture component, resulting in three sets of reference state features. These reference state features are vectorized into 96-dimensional vectors (3 components × 32-dimensional parameters) and stored in the Redis state cache unit in key-value pairs (key: "P1_ref_states", value: 96-dimensional vector), with an expiration time of 24 hours to ensure data timeliness.

[0029] After acquiring the initial reference state features, the server needs to dynamically update the reference state features using real-time data from time point M to time point L (a total of 15 × 60 = 900 sampling points, corresponding to the operating conditions of point P1 from 07:00:00 to 08:15:00, with temperature fluctuations of 170~175℃ and pressure of 78~82 bar). First, the server inputs the real-time time-series data for this period into the thermal feature extraction network to generate 900 32-dimensional real-time thermal state features. Then, for each real-time thermal state feature, the server calculates its Mahalanobis distance (used to measure the thermal state difference between features) with the three sets of reference state features in the state cache unit, finding the target reference state feature with the smallest thermal state difference. Subsequently, the server calculates the adjustment increment based on the deviation between the real-time thermal state characteristics and the target reference state characteristics, combined with a preset state tuning parameter (valued at 0.02, dynamically adjusted according to temperature fluctuations). This update applies the mean vector and covariance matrix of the target reference state characteristics (the mean vector is updated to the sum of the original mean and the adjustment increment, and the covariance matrix is ​​updated to the weighted sum of the original covariance matrix and the adjustment increment). By iterating through all real-time thermal state characteristics from time point M to time point L, the server completes iterative updates to the three sets of reference state characteristics, ultimately obtaining the thermal state characteristics of point P1 at time point L, i.e., the updated three sets of reference state characteristic vectors.

[0030] Determining the predicted temperature control target value based on the temperature synchronization control model:

[0031] After acquiring the thermal state characteristics of point P1, the server combines the process condition parameters of that point, the first control strategy parameters (preset as a temperature control response coefficient of 0.8 and a dynamic compensation weight of 0.2), and the pre-trained temperature synchronization control model to calculate the estimated temperature control target value. This temperature synchronization control model comprises four core modules: a thermal feature extraction network, a state buffer unit, a feature mapping network, and a control model. These modules work together to achieve precise control.

[0032] First, the server inputs the thermal state features (96-dimensional vector) of point P1 into a feature mapping network. This network is a fully connected neural network with two hidden layers (96-dimensional input, 128-dimensional and 64-dimensional hidden layers, 256-dimensional output, and ReLU activation function). Through feature space mapping and feature scale transformation, the thermal state features are aligned to the input scale of the control model, generating 256-dimensional thermal state features after feature space mapping. Next, the server combines this mapped thermal state features, a 32-dimensional process condition parameter vector, the result of the previous round of temperature synchronization control at point P1 (i.e., the output value of the previous control, 178℃, normalized to 0.9), and the first control strategy parameters (vectorized into a 2-dimensional vector) into a 320-dimensional comprehensive input vector, which is then input into the control model (a fully connected neural network with three hidden layers, 320-dimensional input, 256-dimensional, 128-dimensional, and 64-dimensional hidden layers, 1-dimensional output, and ReLU activation function). The control model generates the predicted temperature control target value for point P1 through nonlinear mapping of the comprehensive input vector.

[0033] In this scenario, the server calculates the estimated temperature control target value for point P1 as 181℃ through the above process. This value is within the target temperature range of 180±2℃, which meets the process requirements. The server converts this estimated temperature control target value into a control command (such as adjusting the heating element power to 85%), and sends it to the field PLC system via industrial Ethernet to achieve synchronous temperature control of point P1.

[0034] It should be noted that the aforementioned temperature synchronization control model has been pre-trained using sample data. During training, the server utilizes the time-series data set detected by the first sample mold (including sample mold detection points, temperature control target values, and process condition parameters) to iteratively optimize the feature mapping network and the control factors of the control model, ensuring that the difference between the model's output estimated temperature control target value and the actual target value is minimized. Simultaneously, a lightweight dynamic compensation unit is also included in the model. This unit, while maintaining the control factors of the control model unchanged, corrects the control factors of the feature mapping network and the dynamic compensation unit through error feedback, further improving the model's generalization ability and control accuracy.

[0035] Through the above process, the server achieves real-time monitoring, dynamic feature extraction, and precise control of the temperature at mold detection points, ensuring that the mold temperature remains stable during injection molding and improving product molding quality.

[0036] In this embodiment of the invention, the mold temperature synchronous control model is trained in the following way and can be implemented through the following examples.

[0037] Obtain a first sample mold detection time series data set. The first sample mold detection time series data set includes the sample mold detection point, the temperature control target value of the sample mold detection point at time point M, and the process condition parameters of the sample mold detection point. The time point M is less than or equal to the monitoring period of the sample mold detection point.

[0038] For each sample mold detection time series data, the thermal state characteristics of the sample mold detection points in the sample mold detection time series data are determined. The thermal state characteristics of the sample mold detection points are obtained by updating the state distribution model of at least one reference state feature stored in the state cache unit based on the thermal state characteristics of the sample mold detection points from time point N to time point M. The at least one reference state feature is obtained by performing state distribution modeling on the thermal state characteristics of the sample mold detection points before time point N to obtain N being less than M.

[0039] Based on the thermal state characteristics of the sample mold detection point, the process condition parameters of the sample mold detection point, the first estimated temperature control target value of the sample mold detection point, and the first control strategy parameters, the temperature synchronization control model is trained to obtain a trained temperature synchronization control model. The temperature synchronization control model includes a thermal feature extraction network, the state buffer unit, a priori modeling feature mapping network, and a control model. The first estimated temperature control target value of the sample mold detection point is the result of the previous round of temperature synchronization control of the sample mold detection point at time point M by the temperature synchronization control model.

[0040] The thermal feature extraction network is used to extract the thermal state features corresponding to the detection time series data. The feature mapping network is used to perform feature space mapping and feature scale transformation on the thermal state features, aligning the thermal state features to the input scale of the control model to obtain the thermal state features after feature space mapping. The control model is used to generate the estimated temperature control target value of the sample mold detection point at time point M, taking the thermal state features after feature space mapping, the process condition parameters of the sample mold detection point, the first estimated temperature control target value of the sample mold detection point, and the first control strategy parameters as input.

[0041] In this embodiment of the invention, the training steps are illustrated in detail below using the example of a server training model for synchronous temperature control of an automotive bumper injection mold. The mold contains four cavity detection points (P1~P4). The server constructs a sample set based on production data from the past three months and achieves accurate training by iteratively optimizing the model parameters.

[0042] The server first retrieves the first set of sample mold inspection time-series data from the local database and the MES system. This set contains 1000 sets of sample mold inspection time-series data, covering production records at four points P1 to P4 under different process conditions. Each set of sample data includes: the sample mold inspection point identifier (e.g., "P1" or "P2"), the temperature control target value at time point M (i.e., the actual process requirement temperature at that time point), and the process condition parameters for the sample point. Among them, time point M is the key monitoring point of the sample data, which is shorter than the monitoring period of the sample point (e.g., if the monitoring period for a single sample is 1 hour, time point M is set to the 45th minute within the monitoring period). Taking a set of sample data at point P1 as an example: the sample point is identified as "P1"; the time point M is 2023-09-10 09:45:00, and the target temperature control value at this time point is 180℃ (process requirement for cavity surface temperature); process condition parameters include mold material (H13 steel, code [1,0,0]), injection molding material (ABS resin, melt index 25g / 10min, code [0,1,0]), injection cycle 45s (normalized 0.6), holding pressure 80bar (normalized 0.7), etc., which are vectorized into a 32-dimensional feature vector. All sample data are stored according to point location, and each set of data is accompanied by a timestamp and data quality label (such as "valid" or "abnormal", with abnormal data accounting for <5% and pre-removed). For each set of sample data, the server needs to determine the thermal state characteristics of the sample mold detection points. This characteristic is achieved through two steps: "initial reference state feature modeling - dynamic update". The core is to complete the state distribution modeling and update based on the data before and after time point N (N is less than M, set as the 15th minute within the sample monitoring period, i.e., 30 minutes before M). Taking the above sample data at point P1 as an example, the server first processes the historical data before time point N (2023-09-10 09:15:00): retrieves the time series data of the point N for the previous 7 days (including temperature fluctuation records of 175~185℃ and pressure of 75~85bar), inputs it into the thermal feature extraction network (LSTM network, 128-dimensional input layer, 64-dimensional hidden layer, 32-dimensional output layer, activation function tanh), and extracts a 32-dimensional thermal state feature sequence. Subsequently, the server models the sequence using a Gaussian Mixture Model (GMM), setting three mixture components (corresponding to "low temperature stable state", "heating transition state", and "high temperature stable state"). The EM algorithm is used to estimate the weight, mean vector, and covariance matrix of each component, resulting in three sets of reference state features (vectorized into 96-dimensional vectors), which are then stored in the Redis state cache unit (key: "P1_sample_ref_states").Next, the server updates the reference state features using sample data from time points N to M (09:15:00~09:45:00, a total of 30 minutes, 30×60=1800 sampling points): the time series data of this period (temperature 178~182℃, pressure 78~82bar) is input into the thermal feature extraction network to generate 1800 32-dimensional real-time thermal state features. For each feature, the target reference state feature is matched by calculating the thermal state difference (e.g., Mahalanobis distance) with the reference state features in the cache, and the adjustment increment is calculated in combination with the state tuning parameter (0.02), iteratively updating the mean and covariance of the reference state features. After all features are updated, the server determines the final reference state feature as the thermal state feature (96-dimensional vector) of that sample point. The server trains the model based on the sample thermal state features, process condition parameters, the first predicted temperature control target value (output of the previous training round), and the first control strategy parameters (temperature response coefficient 0.8, dynamic compensation weight 0.2). First, the server inputs the thermal state features (96-dimensional) of sample P1 into a feature mapping network (a 2-layer fully connected network, 96-dimensional input, 128 / 64-dimensional hidden layers, 256-dimensional output, ReLU activation function), and obtains a 256-dimensional mapped feature through spatial mapping and scaling transformation. Then, this mapped feature, a 32-dimensional vector of process condition parameters, the first predicted temperature control target value (179℃ generated in the previous training round, normalized to 0.89), and the first control strategy parameters (vectorized to 2-dimensional) are merged into a 320-dimensional input vector, which is then input into the control model (a 3-layer fully connected network, 320-dimensional input, 256 / 128 / 64-dimensional hidden layers, 1-dimensional output), generating the predicted temperature control target value (e.g., 181℃) for the sample at time point M. The server compares this predicted target value with the actual target value of the sample (180℃), constructs a difference measurement function (e.g., mean squared error), and corrects the control factors of the feature mapping network and the control model through error feedback (using gradient descent to iteratively optimize weights and biases). Repeat the above process, iterating through 1000 sets of sample data, until the average error between the model's estimated target value and the actual target value is less than 0.5℃. At this point, the model training is complete, and the parameters are stored locally on the server (in .h5 files) for subsequent online control. Through this training, the server optimizes the parameters of the temperature synchronization control model, ensuring that the model can accurately output the temperature control target value based on real-time data, meeting the process requirements of mold production.

[0043] In this embodiment of the invention, the determination of the thermal state characteristics of the sample mold detection points in the sample mold detection time series data can be performed through the following examples.

[0044] The sample mold detection points before time point N are imported into the thermal feature extraction network to generate the first thermal state feature.

[0045] The first thermal state feature is modeled into a state distribution using a Gaussian mixture model to obtain at least one reference state feature, and the at least one reference state feature is stored in the state cache unit.

[0046] The sample mold detection points from time point N to time point M are imported into the thermal feature extraction network to generate thermal state features of the sample mold detection points from time point N to time point M.

[0047] Based on the thermal state characteristics of the mold detection points from time point N to time point M, the state distribution model of the at least one reference state feature is updated, and the at least one reference state feature obtained after the update is determined as the thermal state characteristics of the sample mold detection points in the sample mold detection time series data.

[0048] In this embodiment of the invention, taking a set of sample data from point P1 of an automotive bumper injection mold as an example, the process of determining the thermal state characteristics of the sample is described in detail. This sample corresponds to the production period from 09:00:00 to 10:00:00 on September 10, 2023. Time point N is set to 09:15:00 (the 15th minute within the sample monitoring period), and time point M is set to 09:45:00 (the 45th minute within the sample monitoring period, i.e., the key time point for generating the temperature control target value). The server completes the determination of the thermal state characteristics through four steps. First, the server retrieves historical data for point P1 before time point N (09:15:00), specifically the time series data for the 7 days prior to N (including temperature fluctuation records of 170-185℃ and pressure records of 75-85 bar, totaling 7 × 24 × 3600 = 604800 sampling points). The server concatenates these data into a continuous time-series stream in chronological order and inputs it into the thermal feature extraction network (LSTM network, 128-dimensional input layer, 64-dimensional hidden layer, 32-dimensional output layer, activation function tanh) in the temperature synchronization control model. The network models the spatiotemporal correlation of the time-series data, extracting a 32-dimensional first thermal state feature sequence (one feature is generated for every 10 sampling points, totaling 60,480 features). Next, the server models the state distribution of the first thermal state feature sequence: a Gaussian mixture model (GMM) is used to fit the 32-dimensional feature sequence, setting three mixture components (corresponding to the mold's "low-temperature stable state," "heating transition state," and "high-temperature stable state"). The EM algorithm iteratively estimates the weights of each component (e.g., [0.3, 0.4, 0.3]), the mean vector (32-dimensional, e.g., the low-temperature mean [175, 76, ...]), and the covariance matrix (32×32-dimensional), obtaining three sets of reference state features. The server vectorizes these features into 96-dimensional vectors (3 components × 32-dimensional parameters) and stores them in a Redis state cache unit as key-value pairs (key: "P1_sample_ref", value: 96-dimensional vector) to ensure fast retrieval during subsequent updates. Then, the server processes sample data from time points N to M (09:15:00-09:45:00, a total of 30 minutes): it retrieves the time-series data for this period (temperature 178-182℃, pressure 78-82 bar, 30 × 60 = 1800 sampling points), inputs it into the thermal feature extraction network in the same format, and generates 1800 32-dimensional real-time thermal state features (one feature per sampling point). These features reflect the dynamic thermal behavior of the sample points during the N to M time period.Finally, the server updates the reference state features based on the real-time thermal state features from N to M: for each 32-dimensional real-time feature, by calculating its thermal state difference with the three sets of reference state features in the cache (such as Mahalanobis distance, which measures the distribution similarity between features), it matches the target reference state feature with the smallest thermal state difference (e.g., a real-time feature matches "high temperature stable state"); combined with the preset state tuning parameter (0.02, set according to the fluctuation range of the sample data), it calculates the adjustment increment (the deviation between the real-time feature and the target reference state feature × the tuning parameter), and iteratively updates the mean vector and covariance matrix of the target reference state feature (e.g., fine-tuning the mean of "high temperature stable state" from 182℃ to 181.5℃). After traversing 1800 real-time features to complete all updates, the server determines the final three sets of reference state features in the cache (the updated 96-dimensional vector) as the thermal state features of point P1 of the sample, which are used for subsequent model training.

[0049] In this embodiment of the invention, the step of updating the state distribution model of the at least one reference state feature based on the thermal state characteristics of the mold detection points from time point N to time point M can be implemented through the following example.

[0050] Based on the mold detection point characteristics of the sample mold detection point at time point N, the state distribution model of the at least one reference state feature is updated to obtain the updated at least one reference state feature and stored in the state cache unit.

[0051] Based on the mold detection point characteristics at each time point between time point N and time point M, and the thermal state characteristics of the mold detection point at time point M, at least one reference state feature in the state cache unit is updated by state distribution modeling, and at least one updated reference state feature is obtained and stored in the state cache unit.

[0052] In this embodiment of the invention, taking the processing of sample data from point P1 of an automotive bumper injection mold as an example, time point N is 09:15:00 on September 10, 2023 (the start point of the period from N to M), and time point M is 09:45:00 (the end point of the period). The server completes the dynamic update of the reference state features by processing the features of point N and the features of each time point between N and M in stages. First, the server updates the reference state features based on the mold detection point features at time point N. The server extracts the specific data of time point N (09:15:00) from the time series data of the period from N to M: temperature 178.2℃, pressure 78.5 bar, and inputs it into the thermal feature extraction network (LSTM network) to generate 32-dimensional thermal state features of time point N (including information such as temperature fluctuation trend and pressure correlation). Subsequently, the server retrieves three sets of reference state features (low-temperature stable state, warming transition state, and high-temperature stable state, 96-dimensional vector) from the Redis state cache unit. By calculating the thermal state difference between the N features at time point and each set of reference features (using Mahalanobis distance to measure distribution similarity), the server determines the matching target reference state feature. This feature has the smallest difference from the "high-temperature stable state" reference feature (distance value 0.8, less than 1.2 and 1.5 of the other two groups). Based on the preset state tuning parameter (0.02), the server calculates the adjustment increment: increment = 0.02 × (N features at time point - mean vector of target reference feature). The mean vector of the target reference feature (the mean of the high-temperature state in the original 32-dimensional vector [180, 80, ...]) is corrected to obtain the updated mean vector [179.8, 79.9, ...]. At the same time, the covariance matrix is ​​adjusted to adapt to the new feature distribution. After the update, the server revectorizes the adjusted three sets of reference state features into 96-dimensional vectors, overwriting the original "P1_sample_ref" key-value pairs stored in the cache to ensure that the latest reference features are used in subsequent calls. Next, the server continues to update the reference state features based on the features of each time point from N to M and time point M. The N to M time period contains 1800 time points (1 sampling point per second). The server iterates through the features of each time point in chronological order. Taking 09:25:00 (the intermediate time point) as an example, the temperature at this time point is 180.5℃ and the pressure is 80.2 bar. After inputting into the thermal feature extraction network, a 32-dimensional feature is generated. Through thermal difference matching, it is found that this feature has the smallest difference (distance 0.6) from the updated "high-temperature stable state" reference feature. The server then calculates the adjustment increment (0.02 × (current feature - updated mean vector)) to further fine-tune the mean (updated to [180.1, 80.1, ...]) and covariance matrix of this reference feature. For other intermediate time points such as 09:35:00 and 09:44:00, the server repeats the above process, completing the update of features for a total of 1799 intermediate time points.The final processed feature for time point M (09:45:00) is: temperature 181.8℃, pressure 81.0 bar. The generated 32-dimensional feature still matches the "high-temperature stable state" reference feature. The server optimizes the mean vector of this reference feature to [180.3, 80.2, ...] through a final incremental adjustment, and the covariance matrix converges to a stable distribution. The server then stores the three sets of updated reference state features back into the cache unit, completing the state distribution modeling update for the entire time period. Through the above phased updates, the reference state features can adapt to the dynamic thermal behavior of the sample points in real time during the N to M time periods, providing accurate state feature input for subsequent model training.

[0053] In this embodiment of the invention, the mold detection point feature at time point N includes at least two thermal modal features of the detection time series data. The step of updating the state distribution model of the at least one reference state feature based on the mold detection point feature at time point N can be implemented through the following example.

[0054] For the thermal modal characteristics of each detection time series data of the sample mold detection point at time point N, the corresponding target reference state feature is determined. The target reference state feature is the reference state feature with the smallest thermal difference metric between the thermal modal characteristics of the detection time series data.

[0055] Based on the preset state tuning parameters and the thermal modal characteristics of the detection time series data, the target reference state characteristics are modeled and updated for state distribution.

[0056] In this embodiment of the invention, taking the server processing sample data of point P1 of an automotive bumper injection mold as an example, time point N is 09:15:00 on 2023-09-10. The mold detection point features at this time point include two thermal modal features of the detection time series data: "temperature modal features" and "pressure-temperature coupled modal features". The server completes the modeling and updating of the state distribution of the reference state features by processing these two modal features respectively. From the time series data of the period from N to M (09:15:00-09:45:00), the server extracts the original detection data of time point N (09:15:00): temperature 178.2℃, pressure 78.5 bar, and sliding window data of 5 seconds before and after this time point (a total of 11 sampling points to ensure that the features include short-term fluctuation trends). These data are input into a thermal feature extraction network (LSTM network, 128-dimensional input layer, 64-dimensional hidden layer, and 32-dimensional output layer). The network generates 32-dimensional "temperature modal features" (containing information such as temperature change rate and fluctuation amplitude, e.g., the first 10 dimensions of the vector reflect the temperature trend over the past 5 seconds) by modeling the temporal correlation of temperature data. Simultaneously, it generates 32-dimensional "pressure-temperature coupled modal features" (the last 10 dimensions reflect the correlation between pressure and temperature) by modeling the coupling relationship between pressure and temperature (e.g., the lag effect of pressure changes on temperature). These two modal features together constitute the modal detection point features at time point N (64 dimensions in total). The server retrieves three sets of reference state features (low-temperature stable state, heating transition state, and high-temperature stable state, 96-dimensional vectors, 32-dimensional parameters per set) from the Redis state cache unit. It calculates the thermal state differences for each of the two thermal modal features and matches them to the target reference state features. Temperature modal feature matching: The server calculates the thermal state differences between the 32-dimensional temperature modal features and the three sets of reference state features (using Mahalanobis distance to measure the similarity of feature distributions). The calculation results show that the distance to the "high temperature stable state" reference feature is 0.7 (minimum), the distance to the "heating transition state" is 1.3, and the distance to the "low temperature stable state" is 1.6. Therefore, the target reference state feature corresponding to the temperature modal feature is the "high temperature stable state". Pressure-temperature coupled modal feature matching: Similarly, the Mahalanobis distance between the 32-dimensional coupled modal feature and the three sets of reference features is calculated. The results show that the distance to the "high temperature stable state" reference feature is 0.6 (minimum), and the distances to the other two sets are 1.4 and 1.8, respectively. Therefore, the target reference state feature corresponding to the coupled modal feature is also the "high temperature stable state". The server updates the "high temperature stable state" reference feature according to the preset state tuning parameters (0.02, set based on historical data fluctuation amplitude) and the two thermal modal features. Temperature modal feature update: The mean vector of the "high temperature stable state" in the target reference state feature is a 32-dimensional vector μ_high (e.g., [180.0, 80.0, ...], the first 10 dimensions correspond to temperature-related parameters).The server calculates the adjustment increment of the temperature modal feature: increment_temp = 0.02 × (temperature modal feature - μ_high[0:32]), and adds the increment_temp to μ_high to obtain the updated temperature-related mean components [179.8, 79.9, ...]. Coupled modal feature update: The coupled modal feature reflects the correlation between pressure and temperature. The server calculates its adjustment increment: increment_couple = 0.02 × (coupled modal feature - μ_high[0:32]), and further fine-tunes the components in μ_high related to the coupling relationship (such as dimensions 11-20), updating them to [179.8, 79.9, 180.1, ...]. At the same time, the server adjusts the covariance matrix (32 × 32 dimensions) of the "high temperature steady state" reference feature according to the covariance of the two modal features to enhance its adaptability to the current thermal distribution. After the update is complete, the server revectorizes the adjusted three sets of reference state features into 96-dimensional vectors, overwriting the original "P1_sample_ref" key-value pairs in the cache, ensuring that subsequent update steps use the latest reference features. Through the above process, the server accurately locates and updates the target reference state features based on the two thermal modal features at time point N, enabling the reference features to reflect the thermal behavior of sample points at key time points in real time, laying the foundation for dynamic updates in the subsequent N to M time periods.

[0057] In this embodiment of the invention, the step of updating the state distribution model of the target reference state features based on preset state tuning parameters and the thermal modal features of the detection time series data can be implemented through the following example.

[0058] The sum of the target reference state feature and the adjustment increment is determined as the updated target reference state feature. The adjustment increment is the product of the deviation value of the thermal modal feature of the detection time series data after removing the target reference state feature and the state tuning parameter.

[0059] In this embodiment of the invention, taking the sample data of point P1 of an automotive bumper injection mold as an example, the temperature modal feature of time point N (2023-09-10 09:15:00) has matched the target reference state feature as "high temperature stable state". The server updates the target reference state feature according to the following steps: The server retrieves the mean vector μ_high (32-dimensional, such as [180.0, 80.0, 179.5, ...], the first two dimensions corresponding to the core parameters of temperature and pressure) of the "high temperature stable state" reference feature from the Redis state cache unit, and extracts the 32-dimensional temperature modal feature of time point N (such as [178.2, 78.5, 178.0, ...]). Calculate the deviation value: deviation value = temperature modal feature - μ_high, to obtain the 32-dimensional deviation vector (such as [-1.8, -1.5, -1.5, ...]). The server uses a preset state tuning parameter of 0.02 to calculate the adjustment increment: Adjustment increment = Deviation value × 0.02, resulting in a 32-dimensional increment vector (e.g., [-0.036, -0.03, -0.03, ...]). Finally, the server adds the target reference state feature to the adjustment increment to obtain the updated target reference state feature: Updated μ_high = μ_high + Adjustment increment, i.e., [180.0 -0.036, 80.0 -0.03, 179.5 -0.03, ...] = [179.964, 79.97, 179.47, ...]. The server then re-stores the updated "high-temperature stable state" reference feature to the state cache unit, completing the state distribution modeling update.

[0060] In this embodiment of the invention, the step of training the temperature synchronization control model based on the thermal state characteristics of the sample mold detection point, the process condition parameters of the sample mold detection point, the first estimated temperature control target value of the sample mold detection point, and the first control strategy parameters can be implemented through the following example.

[0061] The thermal state features of the sample mold detection points are imported into the feature mapping network to generate thermal state features after feature space mapping.

[0062] The thermal state features mapped by the feature space, the process condition parameters of the sample mold detection point, the first estimated temperature control target value of the sample mold detection point, and the first control strategy parameters are imported into the control model to generate the estimated temperature control target value of the sample mold detection point at time point M.

[0063] Based on the target temperature control value of the sample mold detection point at time point M and the estimated target temperature control value of the sample mold detection point at time point M, the control factors of the feature mapping network and the control model are optimized.

[0064] In this embodiment of the invention, the training process is illustrated in detail below using the example of training a temperature synchronization control model for point P1 of an automotive bumper injection mold. This sample corresponds to production data from 09:45:00 on September 10, 2023 (time point M). The actual temperature control target value is 180℃. The server completes model training in three steps: feature mapping, target value generation, and parameter tuning. The server first retrieves the thermal state features of sample P1. These features are 96-dimensional vectors updated through state distribution modeling (containing parameters such as the mean and covariance of three sets of reference state features, such as the mean vector of the "high-temperature stable state" [179.96, 79.97, ...]). The server inputs these 96-dimensional thermal state features into the feature mapping network (a two-layer fully connected neural network: 96-dimensional input layer, 128 / 64-dimensional hidden layers, and 256-dimensional output layer, with ReLU activation function) in the temperature synchronization control model. The network maps the 96-dimensional reference state distribution information to a 256-dimensional high-dimensional space through nonlinear transformation of thermal state characteristics, while aligning it to the input scale of the control model. For example, the network maps the temperature fluctuation characteristics of the "high-temperature steady state" from the original low-dimensional space to a 256-dimensional vector containing long-term trends and short-term fluctuations, ultimately generating 256-dimensional thermal state characteristics after feature space mapping, which are stored in the server memory for further use. The server retrieves the process condition parameters of sample P1 from the local database (32-dimensional vector, including the mold material H13 steel code [1,0,0], injection cycle 45s normalized to 0.6, holding pressure 80bar normalized to 0.7, etc.), reads the first predicted temperature control target value from the training log (179℃ output from the previous training round, normalized to 0.89 in a 1-dimensional vector), and loads the first control strategy parameters (temperature response coefficient 0.8, dynamic compensation weight 0.2, vectorized into a 2-dimensional vector). These parameters are concatenated with the mapped 256-dimensional thermal state features to form a 320-dimensional comprehensive input vector (256+32+1+2=291 dimensions, normalized to 320 dimensions). The server imports this 320-dimensional vector into the control model (a 3-layer fully connected neural network: 320-dimensional input, 256 / 128 / 64-dimensional hidden layers, 1-dimensional output, ReLU activation function). The model learns the mapping relationship between thermal state features, process parameters, and target temperature values ​​through nonlinear fitting of the comprehensive input vector, outputting the estimated target temperature control value at point P1 at time point M, which, after inverse normalization, is 181℃ (original output 0.92, corresponding to an actual temperature of 181℃). The server extracts the actual target temperature control value of 180℃ at time point M from the sample data, compares it with the estimated target value of 181℃, and calculates the error (1℃). The mean squared error (MSE) is used as the loss function to quantify the bias (loss value = (181-180)² = 1), and the control factors (weight matrix and bias vector) of the feature mapping network and the control model are tuned by gradient descent.For example, in adjusting the connection weights between the output layer and the third hidden layer of the model, if the gradient of a certain weight with respect to the loss is 0.05, the server adjusts this weight by a learning rate of 0.001: new weight = original weight - 0.001 × 0.05, reducing the weight to decrease the error. Similarly, the gradient of the bias vector in the second hidden layer of the feature mapping network is 0.03, and the bias is also adjusted by the learning rate to enhance the accuracy of feature mapping. Forward and backward propagation are iteratively performed until the error between the estimated target value (e.g., 180.3℃) and the actual target value is less than 0.5℃ (loss value 0.09). At this point, the server stores the updated adjustment factors in the model parameter file, completing the training and optimization for this sample and laying the foundation for subsequent training on all samples.

[0065] In this embodiment of the invention, the step of optimizing the control factors of the feature mapping network and the control model based on the target temperature control value of the sample mold detection point at time point M and the estimated target temperature control value of the sample mold detection point at time point M includes:

[0066] Based on the target temperature control value of the sample mold detection point at time point M and the estimated target temperature control value of the sample mold detection point at time point M, a target difference measurement function is constructed.

[0067] Based on the target difference measurement function, the regulation factors of the feature mapping network and the regulation factors of the regulation model are corrected according to error feedback.

[0068] In this embodiment of the invention, taking the training of the P1 point sample of an automotive bumper injection mold as an example, the server, based on the actual and estimated target temperature control values ​​at time point M (2023-09-10 09:45:00), constructs a target difference measurement function and backpropagates the error to optimize the feature mapping network and the control factors of the control model. The server first extracts the actual temperature control target value of 180℃ (the target temperature of the cavity surface required by the process) at point P1 at time point M from the sample data. Combined with the estimated temperature control target value of 181℃ generated by the model, a target difference measurement function is constructed. This function uses mean squared error (MSE) to quantify the deviation between the two values. Specifically, it calculates the squared difference between the actual target value and the estimated target value as the loss value, expressed as "loss value = (actual target value - estimated target value)²". Substituting the data, the loss value is (180-181)² = 1, which reflects the degree of deviation between the model's current output and the actual process requirements. Subsequently, based on the loss value of the target difference metric function, the server corrects the control factors (i.e., the weight matrices and bias vectors of each layer of the network) of the feature mapping network and the control model through an error feedback mechanism. Specifically, the server starts with the loss value and calculates the gradient of the loss function with respect to the output layer of the control model (e.g., the gradient of the weights connecting the output layer and the third hidden layer is 0.05) using the backpropagation algorithm. Then, it backpropagates layer by layer to the input layer of the control model and the output layer of the feature mapping network to determine the contribution of each weight and bias to the loss. For example, the gradient of a certain weight in the third hidden layer of the control model is 0.03, and the gradient of the bias in the second hidden layer of the feature mapping network is 0.02. The server adjusts these control factors according to a preset learning rate (0.001): for weights with a gradient of 0.05, the new weight = original weight - 0.001 × 0.05, reducing the weight to decrease the error; for biases with a gradient of 0.02, the new bias = original bias - 0.001 × 0.02, enhancing the accuracy of the feature mapping. By iteratively executing the aforementioned error feedback and parameter adjustment process (each iteration traverses the sample data, forward propagating to generate the estimated target value, and backpropagating to correct the control factor), the server gradually reduces the loss value. After 5 iterations, the estimated temperature control target value was optimized from 181℃ to 180.3℃, and the error between it and the actual target value of 180℃ was reduced to 0.3℃ (loss value 0.09). At this point, the control factors of the feature mapping network and the control model have been optimized, and the server stores the updated parameters in the model file to ensure that the optimized model is used in the next round of training or online control.

[0069] In this embodiment of the invention, the temperature synchronization control model further includes a lightweight dynamic compensation unit for the model. The temperature synchronization control model is trained based on the thermal state characteristics of the sample mold detection point, the process condition parameters of the sample mold detection point, the first estimated temperature control target value of the sample mold detection point, and the first control strategy parameters. This can be implemented through the following example.

[0070] The thermal state features of the sample mold detection points are imported into the feature mapping network to generate thermal state features after feature space mapping.

[0071] The thermal state features mapped by the feature space, the process condition parameters of the sample mold detection point, the first estimated temperature control target value of the sample mold detection point, and the first control strategy parameters are imported into the control model to generate the estimated temperature control target value of the sample mold detection point at time point M.

[0072] While maintaining the control factor of the control model, the control factors of the feature mapping network and the control factor of the dynamic compensation unit are optimized based on the target temperature control value of the sample mold detection point at time point M and the estimated target temperature control value of the sample mold detection point at time point M.

[0073] In this embodiment of the invention, exemplarily, the following uses server-trained temperature synchronization control model for point P1 of an automotive bumper injection mold as an example. This model adds a lightweight dynamic compensation unit (used for real-time error compensation of the control model output). The training process is as follows: The server retrieves the thermal state features of the P1 sample, and after state distribution modeling and updating, a 96-dimensional vector (containing the mean and covariance parameters of three sets of reference state features, reflecting the dynamic thermal behavior from time N to M) is input into a feature mapping network (a two-layer fully connected neural network: 96-dimensional input, 128 / 64-dimensional hidden layers, 256-dimensional output, with ReLU activation function). The network maps the 96-dimensional thermal state features to a 256-dimensional high-dimensional space through nonlinear transformation, while aligning it to the input scale of the control model, generating thermal state features after feature space mapping (e.g., the first 80 dimensions of the vector enhance temperature stability features, and the last 80 dimensions highlight the pressure-temperature coupling relationship), which are stored in the server memory for further use. The server retrieves the process condition parameters (32-dimensional vector, including mold material H13 steel code [1,0,0], injection cycle 45s normalized to 0.6, etc.), the first estimated temperature control target value (the base value of 179℃ output by the control model in the previous training round, normalized to 0.89), and the first control strategy parameters (temperature response coefficient 0.8, dynamic compensation weight 0.2, vectorized to 2 dimensions) of sample P1 from the local database. These parameters are concatenated with the mapped 256-dimensional thermal state features and merged into a 320-dimensional comprehensive input vector (256+32+1+2=291 dimensions normalized to 320 dimensions). The server imports the 320-dimensional vector into the control model (3-layer fully connected neural network, the control factor remains unchanged from the previous training results, i.e., the weight matrix and bias vector are fixed), and the model outputs the base estimated temperature control target value of 181℃ (original output 0.92, corresponding to 181℃ after inverse normalization). Subsequently, this base value is input into a lightweight dynamic compensation unit (a single-layer fully connected network with 1D input and 1D output, and parameters being the compensation coefficients to be tuned). The unit generates the final estimated temperature control target value of 180.5℃ by superimposing a dynamic compensation value (e.g., -0.5℃, learned based on current process conditions and historical errors) onto the base value. The server extracts the actual temperature control target value of 180℃ (the cavity surface temperature required by the process) at point P1 at time M from the sample data and compares it with the final estimated target value of 180.5℃ after dynamic compensation, calculating the error (0.5℃). The mean squared error (MSE) is used to construct a loss function to quantify the deviation, with a loss value of (180-180.5)²=0.25, reflecting the degree of deviation between the current model output and the true target.The server maintains the control factors of the model (weights and biases unchanged) and corrects the control factors of the feature mapping network and the dynamic compensation unit through error feedback: starting from the loss value, it backpropagates to calculate the gradient of the loss function with respect to the output layer of the dynamic compensation unit (e.g., the gradient of the compensation coefficient is 0.03), and then propagates it to the output layer of the feature mapping network (e.g., the bias gradient of the second hidden layer is 0.02). Based on the preset learning rate (0.001), relevant factors are adjusted: the compensation coefficient of the dynamic compensation unit is finely adjusted from -0.5 to -0.5003 (new coefficient = original coefficient - 0.001 × 0.03), and a certain weight of the second hidden layer of the feature mapping network is adjusted from 0.82 to 0.8198 (new weight = original weight - 0.001 × 0.02). After three rounds of iterative optimization, the predicted target value after dynamic compensation converges to 180.1℃ with an error of 0.1℃ (loss value 0.01). The server stores the updated control factors of the feature mapping network and the dynamic compensation unit in the model file, completing the training.

[0074] In this embodiment of the invention, the optimization process of the control factor of the feature mapping network and the control factor of the dynamic compensation unit based on the target temperature control value of the sample mold detection point at time point M and the estimated target temperature control value of the sample mold detection point at time point M can be implemented through the following example.

[0075] Based on the target temperature control value of the sample mold detection point at time point M and the estimated target temperature control value of the sample mold detection point at time point M, a target difference measurement function is constructed.

[0076] Based on the target difference measurement function, the regulation factors of the feature mapping network and the regulation factors of the dynamic compensation unit are corrected according to the error feedback.

[0077] In this embodiment of the invention, taking the server-trained temperature synchronization control model for point P1 of an automotive bumper injection mold as an example, the model includes a lightweight dynamic compensation unit (used to compensate for errors in the output of the control model). The optimization process is as follows: The server extracts the actual temperature control target value of 180℃ (the target temperature of the cavity surface required by the process) for point P1 at time point M (2023-09-10 09:45:00) from the sample data. Combined with the estimated temperature control target value of 180.5℃ output by the dynamic compensation unit (the basic output of the control model is 181℃, which is reduced by 0.5℃ after correction by the compensation unit), a target difference measurement function is constructed. This function uses mean squared error (MSE) to quantify the deviation between the two. Specifically, it calculates the square difference between the actual target value and the estimated target value as the loss value, expressed as "loss value = (actual target value - estimated target value)²". Substituting the data, the loss value is (180-180.5)² = 0.25. This value reflects the combined effect of the current feature mapping network and the dynamic compensation unit on the model output. Subsequently, based on the loss value of the target difference metric function, the server corrects the adjustment factors (i.e., the network's weight matrix and bias vector) of the feature mapping network and the dynamic compensation unit through an error feedback mechanism. Specifically, the server starts with the loss value and calculates the gradient of the loss function with respect to the output layer of the dynamic compensation unit (the dynamic compensation unit is a single-layer fully connected network, taking a 1-dimensional predicted base value as input and outputting a 1-dimensional compensated target value; its compensation coefficient gradient is 0.03). Then, the error signal is backpropagated to the output layer of the feature mapping network (the gradient of the bias vector in the second hidden layer of the feature mapping network is 0.02, which has the strongest correlation with the dynamic compensation effect). Based on the preset learning rate (0.001), the server optimizes the relevant control factors: the compensation coefficient of the dynamic compensation unit is adjusted from -0.5 (corresponding to a compensation value of -0.5℃) to -0.5003 (new coefficient = original coefficient -0.001 × 0.03), enhancing the accuracy of the correction to the basic output value; a certain weight of the second hidden layer of the feature mapping network is adjusted from 0.82 to 0.8198 (new weight = original weight -0.001 × 0.02), optimizing the mapping effect of thermal state features to the input scale of the control model. By iteratively executing the above error feedback and parameter adjustment (each iteration traverses the sample data forward to generate the estimated target value, and backpropagates to correct the control factors), the server gradually reduces the loss value. After three rounds of iteration, the predicted temperature control target value after dynamic compensation was optimized from 180.5℃ to 180.1℃, and the error with the actual target value of 180℃ was reduced to 0.1℃ (loss value 0.01). At this time, the control factors of the feature mapping network and the dynamic compensation unit have been optimized, and the server stores the updated parameters in the model file to ensure that the model's dynamic compensation capability for process fluctuations meets production requirements.

[0078] In this embodiment of the invention, the following implementation methods are also provided.

[0079] Obtain a second sample mold inspection time sequence data set, wherein the inspection time sequence data of each sample mold in the second sample mold inspection time sequence data set includes the sample mold inspection point and the process condition parameters of the sample mold inspection point;

[0080] By maintaining the control factors of the thermal feature extraction network and the control model, a feature mapping network is trained based on the second sample mold detection time series data set to obtain the prior modeled feature mapping network.

[0081] In this embodiment of the invention, taking the training of the second sample data of the P1 point of the automotive bumper injection mold as an example, the server performs the following steps: The server obtains the second sample mold detection time series data set, containing 2000 sets of samples (covering points P1 to P4, including 3 types of mold materials, 4 types of injection molding materials, and other diverse process conditions, without temperature control target values), and vectorizes the process condition parameters of each sample into 32-dimensional feature vectors. The server loads the trained thermal feature extraction network (LSTM network) and the control model (3-layer fully connected network), and freezes its control factors (weights / biases are fixed). For each set of samples (such as the S136 steel mold and PC material sample at point P1), the time series data is input into the thermal feature extraction network to generate 32-dimensional thermal state features, which are then modeled into 96-dimensional reference state features (3 mixture components) using a Gaussian mixture model (GMM). The server inputs 96-dimensional reference state features into a feature mapping network (a 2-layer fully connected network, 96-dimensional input, 256-dimensional output) to generate mapped thermal state features. These features are then concatenated with 32-dimensional process parameters to form a 320-dimensional vector input to the control model, outputting predicted state indices (such as temperature stability scores). A loss function (mean squared error) is constructed based on the deviation between the sample process condition parameters (e.g., PC material requires a temperature stability score ≥ 0.9) and the predicted state indices. This loss function is then used to correct the control factors (hidden layer weights / biases) of the feature mapping network through backpropagation. After iterative training on 2000 samples for 10 rounds, the adaptability of the 256-dimensional features output by the feature mapping network to the input scale of the control model improves, with the average deviation decreasing from 0.15 to 0.05. The server then stores the network parameters at this point as the "pre-modeled feature mapping network".

[0082] In this embodiment of the invention, the step of training a feature mapping network based on the second sample mold detection time series data set to obtain the prior modeling feature mapping network can be implemented through the following example.

[0083] For each sample mold detection time series data, the sample mold detection points of the sample mold detection time series data are imported into the thermal feature extraction network to generate the thermal state features of the sample mold detection points.

[0084] The thermal state characteristics of the test points of the sample mold are modeled using a Gaussian mixture model to obtain at least one reference state characteristic.

[0085] The at least one reference state feature is input into a feature mapping network to generate thermal state features after feature space mapping.

[0086] The thermal state characteristics mapped by the feature space and the second control strategy parameters are imported into the regulation model to generate the estimated state index of the sample mold detection point.

[0087] Based on the process condition parameters of the sample mold detection points and the estimated state indicators of the sample mold detection points, the control factors of the feature mapping network are optimized to obtain the feature mapping network.

[0088] In this embodiment of the invention, taking the server-trained feature mapping network for point P1 of an automotive bumper injection mold as an example, based on a second sample set (containing 2000 sets of diverse process condition samples), the steps are as follows: The server selects a set of second sample data for point P1 (mold material S136 steel [0,1,0], injection material PC [0,0,1,0], injection cycle 50s normalized to 0.7, holding pressure 90bar normalized to 0.8), and retrieves the time-series data of this sample (3600 sampling points per hour, temperature 175-185℃, pressure 75-85bar). The time-series data is input into the thermal feature extraction network (LSTM network, 128-dimensional input layer, 64-dimensional hidden layer, 32-dimensional output layer, with the control factor frozen). The network generates 3600 32-dimensional thermal state features through time-series modeling (1 feature per 10 sampling points, including temperature fluctuation trend and pressure correlation information). The server models the 32-dimensional thermal state feature sequence using a Gaussian Mixture Model (GMM), setting three mixture components (corresponding to "low-temperature stable state," "heating transition state," and "high-temperature stable state"). The EM algorithm estimates the parameters of each component: weights [0.25, 0.45, 0.3], a 32-dimensional mean vector (e.g., the high-temperature state mean [180, 80, ...]), and a 32×32 covariance matrix. These three sets of parameters are vectorized into a 96-dimensional vector (3 components × 32-dimensional parameters), serving as the reference state features for the sample points and stored in the server's memory. The server inputs these 96-dimensional reference state features into a feature mapping network (a two-layer fully connected network: 96-dimensional input, 128 / 64-dimensional hidden layers, and 256-dimensional output, with ReLU activation). The network uses nonlinear transformations to map the 96-dimensional features to a 256-dimensional high-dimensional space, generating thermal state features after feature space mapping (e.g., the first 80 dimensions enhance temperature stability features, and the last 80 dimensions highlight material adaptability features), aligning and adjusting the model input scale. The server loads the second control strategy parameters (generalization training strategy: stability weight 0.6, material adaptation weight 0.4, vectorized to 2D), concatenates them with 256-dimensional mapping features and 32-dimensional process condition parameters to form a 320-dimensional vector (256+32+2=290 dimensions normalized to 320 dimensions), and inputs it into the control model (3-layer fully connected network, control factors frozen). The model outputs a predicted state index, with a temperature stability score of 0.75 (out of 1.0, reflecting the matching degree between thermal state characteristics and process conditions). The server extracts the requirement from the sample process condition parameters: PC material needs a temperature stability score ≥ 0.9. The calculation bias is 0.15 (0.9-0.75), and a loss function is constructed (mean squared error = 0.15² = 0.0225). The gradients of the feature map network were calculated through backpropagation (output layer weight gradient 0.04, hidden layer bias gradient 0.03), and then optimized with a learning rate of 0.001: the output layer weights were reduced from 0.6 to 0.5996, and the hidden layer biases were reduced from 0.2 to 0.1997.After iterating through 2000 sets of samples for 10 rounds, the average deviation decreased from 0.15 to 0.05. The optimization of the feature mapping network control factor was completed and stored as the parameter file of "preliminary modeling feature mapping network".

[0089] In this embodiment of the invention, determining whether the temperature synchronization control condition has been reached can be performed through the following example.

[0090] If a temperature synchronization control command is received after the timing data stream input of the detection point of the mold to be controlled has ended, it is determined that the temperature synchronization control condition has been met; or...

[0091] If the temperature control is initiated at time point M, and it is determined that the cyclic temperature synchronization control node has been reached, then the temperature synchronization control condition has been reached.

[0092] In this embodiment of the invention, taking the temperature control of point P1 in an automotive bumper injection mold as an example, the server determines whether the temperature synchronization control condition has been reached in the following two ways: The first way is to respond to the control command after the end of the detection timing data stream input. When point P1 completes the production of the 100th bumper injection molded that day (production batch ends), the on-site PLC system stops sending the detection timing data stream (including real-time sampling data of temperature and pressure) to the server and sends a "data stream input ends" signal to the server via industrial Ethernet. After receiving the signal, the server immediately obtains the corresponding temperature synchronization control command from the MES system (command content: "Batch production ends, perform cavity temperature synchronization calibration"), parses the point identifier "P1" and the control type "post-batch calibration" in the command, determines that the trigger condition is met, and then starts the temperature synchronization control process for point P1, retrieving the reference state characteristics from the state cache unit to prepare for thermal state characteristic updates. The second way is to reach the cyclic temperature synchronization control node. The server presets the cyclic control cycle for point P1 to be 30 minutes (based on the thermal inertia characteristics of the mold), and sets 08:00:00 daily as the control start time point M. The server monitors the current time in real time via a local clock module. When the time reaches 08:30:00 (30 minutes after time point M), it automatically checks if the cyclic node has been reached: it retrieves the control start record for point P1 from the Redis cache (the first round of control started at 08:00:00), calculates the difference between the current time and the start time (30 minutes), which matches the preset cyclic cycle. The server determines that the cyclic temperature synchronization control node has been reached and immediately triggers the control process, starting to process the real-time time-series data from time point M to the current node (30 minutes of data from 08:00:00 to 08:30:00, totaling 1800 sampling points), preparing to update the thermal state characteristics and calculate the estimated temperature control target value. Through these two methods, the server can flexibly trigger control based on production process nodes (such as batch end) or a timed cyclic mechanism, ensuring that the mold temperature remains stable throughout the entire production cycle.

[0093] In this embodiment of the invention, the following implementation methods are also provided.

[0094] Input the mold detection points of the mold to be controlled before time point M into the thermal feature extraction network to generate the first thermal state feature.

[0095] The first thermal state feature is modeled using a Gaussian mixture model to obtain at least one reference state feature, and the at least one reference state feature is stored in the state cache unit.

[0096] In this embodiment of the invention, taking the temperature control of point P1 in an automotive bumper injection mold as an example, the server performs the following steps to generate and store reference state features: The server first determines the time point M of the mold detection point P1 to be controlled, and sets the historical data cutoff point before triggering temperature synchronization control as 2023-10-15 07:00:00 (i.e., 1 hour before triggering control). It retrieves historical mold detection point data for point P1 from the local database for 7 days prior to time point M, including time-series data such as temperature (fluctuation between 170-185℃), pressure (75-85 bar), and injection cycle counts, totaling 7 × 24 × 3600 = 604800 sampling points. The server concatenates these data into a continuous time-series stream in chronological order and inputs it into the thermal feature extraction network (LSTM network, 128-dimensional input layer, 64-dimensional hidden layer, 32-dimensional output layer, activation function tanh) in the temperature synchronization control model. Through the network, the spatiotemporal correlation of the time-series data is modeled, and a 32-dimensional first thermal state feature sequence is extracted (one feature is generated for every 10 sampling points, for a total of 60,480 features, including key thermal behavior information such as temperature change rate and pressure coupling trend). Next, the server models the state distribution of the first thermal state feature sequence: a Gaussian mixture model (GMM) is used to fit the 32-dimensional feature sequence, setting three mixture components (corresponding to the mold's "low-temperature stable state," "heating transition state," and "high-temperature stable state"). The EM algorithm is used to iteratively estimate the parameters of each component, the weights [0.3, 0.4, 0.3], the 32-dimensional mean vector (e.g., the low-temperature state mean [175, 76, ...], the high-temperature state mean [180, 80, ...]), and the 32×32-dimensional covariance matrix, resulting in three sets of reference state features. The server vectorizes these features into 96-dimensional vectors (3 components × 32-dimensional parameters), stores them in key-value pairs (key: "P1_ref_states", value: 96-dimensional vector) in the Redis state cache unit, and sets an expiration time of 24 hours to ensure the timeliness of the reference state features, providing an initial benchmark for subsequent real-time data updates of thermal state features based on time points M to L.

[0097] In this embodiment of the invention, the acquisition of the thermal state characteristics of the detection points of the mold to be controlled can be performed through the following example.

[0098] The mold detection points from time point M to time point L are imported into the thermal feature extraction network to generate thermal state features of the mold detection points from time point M to time point L.

[0099] Based on the thermal state characteristics of the mold detection points from time point M to time point L, the state distribution model of the at least one reference state feature is updated, and the at least one reference state feature obtained after the update is determined as the thermal state characteristics of the mold detection points to be controlled.

[0100] In this embodiment of the invention, taking the temperature control of point P1 of an automotive bumper injection mold as an example, the server obtains its thermal state characteristics according to the following steps: The server first determines the time point M (the historical data cutoff point before triggering control, set as 07:00:00 on 2023-10-15) and the time point L (the current monitoring period for identifying the point to be controlled, i.e., the trigger time 08:15:00). The period from M to L is 1 hour and 15 minutes (75 minutes), corresponding to 75×60=4500 sampling points (1 sampling point per second). The server retrieves the time series data of the mold detection point during this period from the local database: temperature fluctuates between 170-175℃ (average 172.5℃), pressure is 78-82 bar (average 80 bar), including the real-time monitoring values ​​of the temperature sensor and the pressure sensor. These time-series data are concatenated into a continuous data stream in chronological order and input into the thermal feature extraction network (LSTM network, 128-dimensional input layer, 64-dimensional hidden layer, 32-dimensional output layer, activation function tanh) in the temperature synchronization control model. The network generates 4500 32-dimensional thermal state features (one feature per sampling point) by modeling the spatiotemporal correlations of the time-series data (e.g., temperature change rate, hysteresis correlation between pressure and temperature). These features reflect the dynamic thermal behavior of point P1 during the M to L period (e.g., low-temperature fluctuation trend, slow temperature decrease under stable pressure). Next, the server updates the reference state features based on the thermal state features during the M to L period. Three sets of reference state features (low-temperature stable state, heating transition state, high-temperature stable state, 96-dimensional vectors, each containing weights, mean vectors, and covariance matrices) are retrieved from the Redis state cache unit. For 4500 32-dimensional real-time thermal state features, the server calculates the thermal state difference between each feature and three sets of reference features (using Mahalanobis distance to measure distribution similarity), and matches the target reference state features. Calculations show that 92% of the real-time features have the smallest difference from the "low-temperature stable state" reference features (distance value 0.9-1.2, less than the other two groups' 1.5-2.0). Based on a preset state optimization parameter (0.02, set according to historical data fluctuation amplitude), the server calculates the adjustment increment: increment = 0.02 × (real-time feature - target reference feature mean vector). The mean vector of the "low-temperature stable state" reference features (the temperature-related components [175,76,...] in the original 32-dimensional vector) is iteratively corrected, gradually adjusted to [173,77,...]. Simultaneously, the covariance matrix is ​​updated to adapt to the feature distribution from time M to L (e.g., the temperature fluctuation variance increases from 0.8 to 1.2). After updating all 4,500 real-time features, the server determines the final three sets of reference state features (updated 96-dimensional vectors) as the thermal state features of point P1 and stores them in local memory for subsequent use by the temperature control model.

[0101] In this embodiment of the invention, the step of determining the estimated temperature control target value of the mold detection point based on the thermal state characteristics of the mold detection point to be controlled, the process condition parameters of the mold detection point to be controlled, the first control strategy parameters, and the pre-trained temperature synchronization control model can be implemented through the following example.

[0102] The thermal state features of the detection points of the mold to be controlled are imported into the feature mapping network to generate thermal state features after feature space mapping.

[0103] The thermal state characteristics mapped by the feature space, the process condition parameters of the detection point of the mold to be regulated, the result of the previous round of temperature synchronization regulation of the detection point of the mold to be regulated, and the first control strategy parameters are imported into the regulation model to generate the estimated temperature regulation target value of the detection point of the mold to be regulated.

[0104] In this embodiment of the invention, taking the temperature control of point P1 in an automotive bumper injection mold as an example, the server determines the estimated temperature control target value according to the following steps: The server retrieves the thermal state characteristics of point P1, and updates the resulting 96-dimensional vector (containing the mean and covariance parameters of three sets of reference state characteristics, reflecting the dynamic thermal behavior from M to L: low temperature steady state mean [173,77,...], weight 0.6; heating transition state mean [176,79,...], weight 0.3; high temperature steady state mean [180,80,...], weight 0.1) into the feature mapping network (2-layer fully connected neural network: 96-dimensional input, 128-dimensional first hidden layer, 64-dimensional second hidden layer, 256-dimensional output, activation function ReLU) in the temperature synchronization control model. The network maps 96-dimensional thermal state features to a 256-dimensional high-dimensional space through nonlinear transformations (such as ReLU activation in the first hidden layer to enhance temperature fluctuation features and ReLU activation in the second hidden layer to highlight the pressure-temperature coupling relationship). This generates thermal state features after feature space mapping (such as the first 100 dimensions of the vector to enhance low-temperature stability features and the last 100 dimensions to highlight material adaptability features). The network aligns and controls the input scale of the model and stores it in the server memory for further use. The server retrieves the process condition parameters for point P1 from the local database: a 32-dimensional feature vector (mold material H13 steel code [1,0,0], cavity size normalized [0.9,0.7,0.5], injection molding material ABS resin code [0,1,0], injection cycle 45s normalized 0.6, holding pressure 80bar normalized 0.7, etc.); it reads the result of the previous round of temperature synchronization control from the control log: 178℃ (the temperature value output by the previous control, normalized to a 1-dimensional vector of 0.89); and loads the first control strategy parameters: temperature control response coefficient 0.8, dynamic compensation weight 0.2 (vectorized into a 2-dimensional vector). The above parameters are concatenated with the mapped 256-dimensional thermal state features to form a 320-dimensional comprehensive input vector (256+32+1+2=291 dimensions normalized to 320 dimensions). The server imports a 320-dimensional vector into the control model (a 3-layer fully connected neural network: 320-dimensional input, 256-dimensional first hidden layer, 128-dimensional second hidden layer, 64-dimensional third hidden layer, and 1-dimensional output, with ReLU activation function). The model, through nonlinear fitting of the comprehensive input vector (e.g., learning the low-temperature trend in thermal state characteristics and the mapping relationship between material melting characteristics and temperature target values ​​in process parameters), outputs a predicted temperature control target value for point P1, which, after inverse normalization, is 181℃ (original output 0.92, corresponding to an actual temperature of 181℃, within the process requirement of 180±2℃). The server stores this predicted temperature control target value in its local control instruction library, preparing to generate heating tube power adjustment instructions (e.g., increasing power from the current 60% to 85%) to achieve synchronous temperature control of point P1.

[0105] This invention provides a computer device 100, which includes a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device 100 executes the aforementioned mold temperature synchronization control method. Figure 2 As shown, Figure 2 This is a structural block diagram of a computer device 100 provided in an embodiment of the present invention. The computer device 100 includes a memory 111, a processor 112, and a communication unit 113. To enable data transmission or interaction, the memory 111, processor 112, and communication unit 113 are electrically connected to each other directly or indirectly. For example, these components can be electrically connected to each other through one or more communication buses or signal lines.

[0106] For illustrative purposes, the foregoing description has been made with reference to specific embodiments. However, the foregoing illustrative discussions are not intended to be exhaustive or to limit the present disclosure to the precise forms disclosed. Numerous modifications and variations are possible in accordance with the foregoing teachings. These embodiments were chosen and described in order to best illustrate the principles of the present disclosure and its practical application, thereby enabling those skilled in the art to best utilize the disclosure and to employ various embodiments with different modifications to suit a particular intended application.

Claims

1. A method for synchronously controlling mold temperature, characterized in that, include: The detection points of the mold to be adjusted and the process condition parameters of the detection points of the mold to be adjusted are obtained, and the detection points of the mold to be adjusted are accessed in a standardized manner using the detection timing data stream; In response to reaching the temperature synchronization control condition, the thermal state characteristics of the detection point of the mold to be controlled are acquired. The thermal state characteristics of the detection point of the mold to be controlled are obtained by updating the state distribution model of at least one reference state characteristic stored in the state cache unit based on the thermal state characteristics of the detection point of the mold to be controlled from time point M to time point L. The at least one reference state characteristic is obtained by updating the state distribution model of the thermal state characteristics of the detection point of the mold to be controlled before time point M. The time point L is the monitoring period of the detection point of the mold to be controlled when the detection point of the mold to be controlled is identified. The L is less than or equal to the monitoring period of the detection point of the mold to be controlled. Based on the thermal state characteristics of the detection point of the mold to be regulated, the process condition parameters of the detection point of the mold to be regulated, the first control strategy parameters, and the pre-trained temperature synchronization regulation model, the estimated temperature regulation target value of the detection point of the mold to be regulated is determined. The temperature synchronization regulation model includes a thermal feature extraction network, the state buffer unit, a feature mapping network, and a regulation model. The thermal feature extraction network is used to extract the thermal state features corresponding to the detection time series data. The feature mapping network is used to perform feature space mapping and feature scale transformation on the thermal state features, aligning the thermal state features to the input scale of the control model to obtain the thermal state features after feature space mapping. The control model is used to generate the estimated temperature control target value of the detection point of the mold to be controlled by taking the thermal state features after feature space mapping, the process condition parameters of the detection point of the mold to be controlled, the result of the previous round of temperature synchronization control of the detection point of the mold to be controlled, and the first control strategy parameters as input.

2. The method according to claim 1, characterized in that, The temperature synchronization control model is trained in the following ways: Obtain a first sample mold detection time series data set. The first sample mold detection time series data set includes the sample mold detection point, the temperature control target value of the sample mold detection point at time point M, and the process condition parameters of the sample mold detection point. The time point M is less than or equal to the monitoring period of the sample mold detection point. For each sample mold detection time series data, the thermal state characteristics of the sample mold detection points in the sample mold detection time series data are determined. The thermal state characteristics of the sample mold detection points are obtained by updating the state distribution model of at least one reference state feature stored in the state cache unit based on the thermal state characteristics of the sample mold detection points from time point N to time point M. The at least one reference state feature is obtained by performing state distribution modeling on the thermal state characteristics of the sample mold detection points before time point N to obtain N being less than M. Based on the thermal state characteristics of the sample mold detection point, the process condition parameters of the sample mold detection point, the first estimated temperature control target value of the sample mold detection point, and the first control strategy parameters, the temperature synchronization control model is trained to obtain a trained temperature synchronization control model. The temperature synchronization control model includes a thermal feature extraction network, the state buffer unit, a priori modeling feature mapping network, and a control model. The first estimated temperature control target value of the sample mold detection point is the result of the previous round of temperature synchronization control of the sample mold detection point at time point M by the temperature synchronization control model. The thermal feature extraction network is used to extract the thermal state features corresponding to the detection time series data. The feature mapping network is used to perform feature space mapping and feature scale transformation on the thermal state features, aligning the thermal state features to the input scale of the control model to obtain the thermal state features after feature space mapping. The control model is used to generate the estimated temperature control target value of the sample mold detection point at time point M, taking the thermal state features after feature space mapping, the process condition parameters of the sample mold detection point, the first estimated temperature control target value of the sample mold detection point, and the first control strategy parameters as input.

3. The method according to claim 2, characterized in that, The determination of the thermal state characteristics of the sample mold detection points in the sample mold detection time series data includes: The sample mold detection points before time point N are imported into the thermal feature extraction network to generate the first thermal state feature. The first thermal state feature is modeled into a state distribution using a Gaussian mixture model to obtain at least one reference state feature, and the at least one reference state feature is stored in the state cache unit. The sample mold detection points from time point N to time point M are imported into the thermal feature extraction network to generate thermal state features of the sample mold detection points from time point N to time point M. For the thermal modal characteristics of each detection time series data of the sample mold detection point at time point N, the corresponding target reference state feature is determined. The target reference state feature is the reference state feature with the smallest thermal difference metric between the thermal modal characteristics of the detection time series data. The sum of the target reference state feature and the adjustment increment is determined as the updated target reference state feature. The adjustment increment is the product of the thermal modal feature of the detection time series data, the deviation value of the target reference state feature, and the state tuning parameter. At least one updated reference state feature is obtained and stored in the state cache unit. Based on the mold detection point characteristics at each time point between time point N and time point M, and the thermal state characteristics of the mold detection point at time point M, at least one reference state feature in the state cache unit is updated by state distribution modeling. The updated at least one reference state feature is obtained and stored in the state cache unit. The updated at least one reference state feature is determined as the thermal state characteristics of the sample mold detection point in the sample mold detection time series data.

4. The method according to claim 2, characterized in that, The step of training a temperature synchronization control model based on the thermal state characteristics of the sample mold detection points, the process condition parameters of the sample mold detection points, the first estimated temperature control target value of the sample mold detection points, and the first control strategy parameters includes: The thermal state features of the sample mold detection points are imported into the feature mapping network to generate thermal state features after feature space mapping. The thermal state features mapped by the feature space, the process condition parameters of the sample mold detection point, the first estimated temperature control target value of the sample mold detection point, and the first control strategy parameters are imported into the control model to generate the estimated temperature control target value of the sample mold detection point at time point M. Based on the target temperature control value of the sample mold detection point at time point M and the estimated target temperature control value of the sample mold detection point at time point M, a target difference measurement function is constructed. Based on the target difference measurement function, the regulation factors of the feature mapping network and the regulation factors of the regulation model are corrected according to the error feedback.

5. The method according to claim 2, characterized in that, The temperature synchronization control model also includes a lightweight dynamic compensation unit. The training of the temperature synchronization control model based on the thermal state characteristics of the sample mold detection points, the process condition parameters of the sample mold detection points, the first estimated temperature control target value of the sample mold detection points, and the first control strategy parameters includes: The thermal state features of the sample mold detection points are imported into the feature mapping network to generate thermal state features after feature space mapping. The thermal state features mapped by the feature space, the process condition parameters of the sample mold detection point, the first estimated temperature control target value of the sample mold detection point, and the first control strategy parameters are imported into the control model to generate the estimated temperature control target value of the sample mold detection point at time point M. By maintaining the control factor of the control model, and based on the target temperature control value of the sample mold detection point at time point M and the estimated target temperature control value of the sample mold detection point at time point M, a target difference measurement function is constructed. Based on the target difference measurement function, the regulation factors of the feature mapping network and the regulation factors of the dynamic compensation unit are corrected according to the error feedback.

6. The method according to claim 2, characterized in that, The method further includes: Obtain a second sample mold inspection time sequence data set, wherein the inspection time sequence data of each sample mold in the second sample mold inspection time sequence data set includes the sample mold inspection point and the process condition parameters of the sample mold inspection point; While maintaining the control factors of the thermal feature extraction network and the control model, for each sample mold detection time series data, the sample mold detection points of the sample mold detection time series data are imported into the thermal feature extraction network to generate the thermal state features of the sample mold detection points. The thermal state characteristics of the test points of the sample mold are modeled using a Gaussian mixture model to obtain at least one reference state characteristic. The at least one reference state feature is input into a feature mapping network to generate thermal state features after feature space mapping. The thermal state characteristics mapped by the feature space and the second control strategy parameters are imported into the regulation model to generate the estimated state index of the sample mold detection point. Based on the process condition parameters of the sample mold detection points and the estimated state indicators of the sample mold detection points, the control factors of the feature mapping network are optimized to obtain the feature mapping network.

7. The method according to claim 1, characterized in that, The method further includes: Input the mold detection points of the mold to be controlled before time point M into the thermal feature extraction network to generate the first thermal state feature. The first thermal state feature is modeled using a Gaussian mixture model to obtain at least one reference state feature, and the at least one reference state feature is stored in the state cache unit.

8. The method according to claim 1, characterized in that, The process of acquiring the thermal state characteristics of the detection points of the mold to be controlled includes: The mold detection points from time point M to time point L are imported into the thermal feature extraction network to generate thermal state features of the mold detection points from time point M to time point L. Based on the thermal state characteristics of the mold detection points from time point M to time point L, the state distribution model of the at least one reference state feature is updated, and the at least one reference state feature obtained after the update is determined as the thermal state characteristics of the mold detection points to be controlled.

9. The method according to claim 1, characterized in that, The step of determining the estimated temperature control target value of the mold detection point based on the thermal state characteristics of the detection point, the process condition parameters of the detection point, the first control strategy parameters, and the pre-trained temperature synchronization control model includes: The thermal state features of the detection points of the mold to be controlled are imported into the feature mapping network to generate thermal state features after feature space mapping. The thermal state characteristics mapped by the feature space, the process condition parameters of the detection point of the mold to be regulated, the result of the previous round of temperature synchronization regulation of the detection point of the mold to be regulated, and the first control strategy parameters are imported into the regulation model to generate the estimated temperature regulation target value of the detection point of the mold to be regulated.

10. A server system, characterized in that, Includes a server, the server being used to perform the method according to any one of claims 1-9.

Images