Production control method and system for phase change material

Abstract

The application provides a production control method and system of a phase change material. The method comprises: obtaining multimodal data of a mixed liquid during stirring when stirring duration of the mixed liquid reaches a set duration; extracting spatiotemporal correlation features from reflective spectral features, liquid surface fluctuation amplitudes and bubble distribution features by using a spatiotemporal convolutional neural network; extracting apparent viscosity time series features from torque feedback data by using a rheological feature extraction module; inputting the spatiotemporal correlation features and the apparent viscosity time series features into a gated attention mechanism for dynamic weighting to obtain a multimodal representation vector; inputting the multimodal representation vector into an improved Transformer time series classifier to dynamically output a stirring state judgment result based on a progressive classification threshold; and triggering a process switching signal when the judgment reaches a dense stable state. The application can significantly improve the accuracy and stability of dense stable state recognition.

Description

Technical Field

[0001] This invention relates to the field of production control technology, and more specifically, to a production control method and system for phase change materials. Background Technology

[0002] Phase change materials (PCMs) are functional materials that can absorb or release a large amount of latent heat through a phase change process within a specific temperature range. Cooling particle PCMs, as a type of PCM, can provide a comfortable cooling experience in clothing, home furnishings, and other fields. Their preparation process involves emulsifying octadecane with a paraffin emulsifier, followed by multi-step mixing and granulation with raw materials such as nano-silica powder, polyurethane solution, and nano-aluminum powder. This process requires high precision in production control.

[0003] In the preparation of cool-feeling particle phase change materials, steps include adding nano-silica powder to an emulsified paraffin solution and stirring, followed by adding a polyurethane solution until a stable, viscous liquid is formed. The viscous, stable state of the mixture is a key indicator of whether the stirring is sufficient and whether it can proceed to the next step. However, existing technologies struggle to identify the viscous, stable state using computer vision during stirring. This is because the consistency change during stirring is a continuous and complex process. Different raw material ratios and stirring rates result in variations in the appearance of the viscous, stable state, such as color depth, fluidity changes, and bubble distribution. Computer vision systems struggle to accurately capture these subtle features and establish effective recognition models. Often, manual observation and judgment are required, which is not only highly subjective and prone to inconsistencies due to operator experience, but also lags behind the actual reaction process, affecting production efficiency and product quality stability.

[0004] Therefore, there is an urgent need for a production control method for phase change materials that can accurately identify the state of the mixture during the stirring process, so as to improve the stability of product performance and reduce production costs. Summary of the Invention

[0005] In order to solve the technical problems existing in the background art, the present invention provides a production control method, system, electronic device, computer storage medium and computer program product for phase change materials.

[0006] The first aspect of the present invention provides a method for controlling the production of phase change materials, comprising the following steps:

[0007] When the stirring time of the mixed liquid reaches the set time, multimodal data of the mixed liquid during the stirring process are acquired, including reflectance spectral characteristics, liquid surface fluctuation amplitude, bubble distribution characteristics, and torque feedback data.

[0008] Spatiotemporal convolutional neural networks are used to extract spatiotemporal correlation features from reflectance spectrum features, liquid surface fluctuation amplitude, and bubble distribution features. A rheological feature extraction module is used to extract apparent viscosity time-series features from torque feedback data. The spatiotemporal correlation features and apparent viscosity time-series features are input into a gating attention mechanism for dynamic weighting to obtain a multimodal representation vector.

[0009] The multimodal representation vector is input into an improved Transformer temporal classifier, which dynamically outputs the stirring state determination result based on a progressive classification threshold. When the determination reaches a viscous stable state, a process switching signal is triggered. The temporal classifier introduces relative position encoding to enhance the correlation of local features.

[0010] A second aspect of the present invention provides a production control system for phase change materials, the system comprising an acquisition unit, a feature extraction unit, and a stability determination unit;

[0011] The acquisition unit acquires multimodal data of the mixed liquid during the stirring process when the stirring time of the mixed liquid reaches the set time, including reflectance spectral characteristics, liquid surface fluctuation amplitude, bubble distribution characteristics, and torque feedback data.

[0012] The feature extraction unit uses a spatiotemporal convolutional neural network to extract spatiotemporal correlation features from reflectance spectrum features, liquid surface fluctuation amplitude, and bubble distribution features. It also uses a rheological feature extraction module to extract apparent viscosity time-series features from torque feedback data. The spatiotemporal correlation features and apparent viscosity time-series features are input into a gating attention mechanism for dynamic weighting to obtain a multimodal representation vector.

[0013] The stability determination unit inputs the multimodal representation vector into an improved Transformer temporal classifier and dynamically outputs the stirring state determination result based on a progressive classification threshold. When the determination reaches a viscous stable state, a process switching signal is triggered. The temporal classifier introduces relative position encoding to enhance the correlation of local features.

[0014] A third aspect of the present invention provides an electronic device comprising: a memory storing executable program code; a processor coupled to the memory; the processor invoking the executable program code stored in the memory to perform the method as described in any of the preceding claims.

[0015] A fourth aspect of the present invention provides a computer storage medium storing a computer program that, when executed by a processor, performs the method described in any of the preceding claims.

[0016] A fifth aspect of the invention provides a computer program product comprising a computer program for performing the method as described in any of the preceding claims.

[0017] This invention, through multimodal data fusion and spatiotemporal feature extraction, combined with an improved Transformer classifier and dynamic threshold determination, can accurately identify the viscous stable state of mixed liquids. Compared with traditional manual judgment or single-parameter control methods, it significantly improves the accuracy and stability of viscous stable state identification, reduces quality fluctuations caused by subjective errors, enables automatic process switching, and significantly improves production efficiency. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic flowchart of a production control method for phase change materials disclosed in an embodiment of the present invention;

[0020] Figure 2 This is a schematic flowchart of a production control system for phase change materials disclosed in an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, exemplary embodiments will be described in detail below, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0022] The production process of phase change materials is roughly as follows: Octadecylane is emulsified using a paraffin emulsifier to obtain a well-emulsified paraffin solution; nano-silica powder, polyurethane solution, and nano-aluminum powder are provided; the nano-silica powder is added to the well-emulsified paraffin solution and stirred evenly to obtain a first mixed liquid; the polyurethane solution is added to the first mixed liquid until a stable, viscous second mixed liquid is formed; and the nano-aluminum powder is added to the second mixed liquid to obtain a third mixed liquid, which is then added to a granulator for granulation to obtain cool-feeling phase change material particles. In this invention, the mixed liquid refers to the second mixed liquid.

[0023] like Figure 1As shown, an embodiment of the present invention provides a method for controlling the production of a phase change material, comprising the following steps:

[0024] S1, when the stirring time of the mixed liquid reaches the set time, acquires multimodal data of the mixed liquid during the stirring process, including reflectance spectral characteristics, liquid surface fluctuation amplitude, bubble distribution characteristics, and torque feedback data.

[0025] In this step, when the stirring time of the mixed liquid reaches the set time, it can be preliminarily determined that it has reached or is close to a viscous stable state. At this time, the data acquisition process is initiated to obtain multimodal data of the mixed liquid during the stirring process. Specifically, the reflectance spectral characteristics of the mixed liquid are obtained through a hyperspectral imager, the amplitude of liquid surface fluctuations is monitored using a laser displacement sensor array, the bubble distribution characteristics are quantified by a polarization vision module, and the torque feedback data of the embedded rheometer is collected simultaneously.

[0026] S2 uses a spatiotemporal convolutional neural network to extract spatiotemporal correlation features from reflectance spectrum features, liquid surface fluctuation amplitude, and bubble distribution features. It also uses a rheological feature extraction module to extract apparent viscosity time-series features from torque feedback data. The spatiotemporal correlation features and apparent viscosity time-series features are input into a gating attention mechanism for dynamic weighting to obtain a multimodal representation vector.

[0027] First, a spatiotemporal convolutional neural network is used to process the reflectance spectrum features, liquid surface fluctuation amplitude, and bubble distribution features. This spatiotemporal convolutional neural network can analyze the above data in both time and space dimensions, extract the spatiotemporal correlation features between them, and capture the dynamic relationship between different features over time and their spatial mutual influence.

[0028] Simultaneously, the apparent viscosity time-series features are extracted from the torque feedback data using a rheological feature extraction module. Torque feedback data is directly related to the apparent viscosity of the liquid; this rheological feature extraction module reveals the viscosity variation over time, reflecting the changes in the rheological properties of the mixed liquid during stirring. The specific processing procedure is as follows:

[0029] (1) Torque signal preprocessing

[0030] The Savitzky-Golay smoothing filter algorithm is used to process the original torque feedback data. While preserving the signal trend characteristics, it eliminates high-frequency noise such as mechanical vibration of the stirring equipment and electromagnetic interference from the sensors, making the signal curve smoother. A linear detrending method is used to remove baseline drift in the torque signal (such as slow trend shifts caused by equipment temperature changes), ensuring that the data reflects the true resistance changes of the mixed liquid.

[0031] The corrected torque data is Z-score standardized to convert the data into a distribution with a mean of 0 and a standard deviation of 1, thus eliminating the influence of differences in the initial state of different batches of raw materials on subsequent calculations.

[0032] (2) Torque-viscosity conversion (based on rheological principles)

[0033] Calculate the specific geometric constant K of the equipment based on the geometric parameters of the agitator (such as blade radius, impeller area, and agitator shaft speed), and then use the formula... , to the torque value Shear stress, which is converted into the force that resists deformation within the liquid. Combining the stirring speed (rpm) and the stirrer radius, an empirical formula is used... ( The coefficients are related to the equipment structure. Calculate the shear rate (for rotational speed). (Reflects the rate at which a liquid is sheared and deformed). Based on fundamental rheological relationships. The ratio of shear stress to shear rate is used as the apparent viscosity value. This enables the physical conversion from torque signals to viscosity parameters.

[0034] (3) Temporal feature extraction (sliding window analysis)

[0035] Continuous apparent viscosity data are divided into overlapping time windows with fixed window lengths (e.g., 50 sampling points) and step sizes (e.g., 10 sampling points), each window corresponding to a local stage of the stirring process. Basic statistics are calculated for the viscosity data within each window, including mean (reflecting the average viscosity level of that stage), standard deviation (reflecting the degree of viscosity fluctuation), skewness, and kurtosis (reflecting the symmetry and steepness of the viscosity distribution).

[0036] The slope of the viscosity data within a window is calculated using linear fitting to characterize the upward / downward trend of viscosity during that stage; the rate of viscosity change is captured by calculating the difference between the means of adjacent windows. A Fast Fourier Transform (FFT) is performed on the data within the window to extract the dominant frequency component (reflecting the periodic characteristics of viscosity fluctuations), which is used to identify stable / unstable states during the stirring process.

[0037] (4) Dynamic rheological feature modeling

[0038] A bidirectional long short-term memory (LSTM) network is used to model the window feature sequence. The forward and backward hidden layers capture the historical dependence of viscosity changes (such as the influence of previous viscosity on the current state) and future trends (such as upcoming viscosity mutations).

[0039] Meanwhile, an attention mechanism is introduced into the LSTM output layer to assign dynamic weights to features in different time windows. For example, when viscosity suddenly increases or decreases, the weight of the corresponding window is automatically increased, strengthening the contribution of key change points to the overall features. The weighted temporal features are then nonlinearly transformed using a fully connected neural network, compressing the high-dimensional features into fixed-dimensional vectors, ultimately forming apparent viscosity temporal features that reflect the dynamic changes in viscosity throughout the stirring process.

[0040] Subsequently, spatiotemporal correlation features and apparent viscosity time-series features are input into a gated attention mechanism. The gated attention mechanism dynamically weights features according to their importance, giving higher weights to features that are more critical to determining the state of the mixed liquid. Finally, these features are fused to generate a multimodal representation vector, achieving effective integration of different types of features.

[0041] S3, the multimodal representation vector is input into the improved Transformer temporal classifier, and the stirring state determination result is dynamically output based on the progressive classification threshold. When the determination reaches the viscous stable state, the process switching signal is triggered; wherein, the temporal classifier introduces relative position encoding to enhance the correlation of local features.

[0042] The improved Transformer temporal classifier introduces relative position encoding, which enhances the capture of local feature correlations to better understand the relationships between local features in the multimodal representation vector and improve classification accuracy.

[0043] The time-series classifier dynamically outputs the stirring state determination result based on a progressive classification threshold. This progressive classification threshold adjusts according to different stages of the stirring process, adapting to the dynamic changes in the mixed liquid's state and more accurately determining whether a viscous stable state has been reached. When the mixed liquid is determined to have reached a viscous stable state, a process switching signal is triggered, instructing the production process to proceed to the next step, achieving automated control of the production process and improving production efficiency and accuracy.

[0044] This invention, through multimodal data fusion and spatiotemporal feature extraction, combined with an improved Transformer classifier and dynamic threshold determination, can accurately identify the viscous stable state of mixed liquids. Compared with traditional manual judgment or single-parameter control methods, it significantly improves the accuracy and stability of viscous stable state identification, reduces quality fluctuations caused by subjective errors, enables automatic process switching, and significantly improves production efficiency.

[0045] Optionally, the set duration is determined in the following manner:

[0046] The first theoretical time required for the reaction system to reach a uniform mixing state is calculated based on the raw material ratio data of the mixed liquid. The second theoretical time required for the fluid to reach a uniform mixing state is calculated by combining the geometric parameters of the stirrer and the fluid dynamics formula. The baseline time is calculated by weighting the first theoretical time and the second theoretical time.

[0047] By analyzing the viscosity change curve in historical production data, the duration during which the viscosity growth rate drops below the first percentage of the initial stage is determined. Based on this duration, the baseline duration is corrected to obtain the set duration.

[0048] Based on the raw material ratio data of the mixed liquid (such as the mass percentages of octadecane, nano-silica powder, and polyurethane solution), the first theoretical time for the reaction system to reach a homogeneous mixing state is calculated using a reaction kinetic model (such as a reaction rate equation based on concentration changes). For example, when the proportion of polyurethane solution increases, its crosslinking reaction with the paraffin emulsion is delayed, and the first theoretical time increases accordingly.

[0049] Meanwhile, by combining the geometric parameters of the agitator (such as blade radius, number of blades, and agitator shaft speed), the second theoretical time required for the fluid to achieve uniform mixing is calculated using fluid dynamics formulas (such as the correlation between Reynolds number and mixing time), reflecting the impact of the equipment on mixing efficiency.

[0050] The first theoretical duration and the second theoretical duration are weighted and calculated (e.g., the raw material reaction characteristics are weighted at 60% and the equipment parameters at 40%) to obtain the initial baseline duration.

[0051] Next, analyze the viscosity change curves in historical production data to determine the actual time it takes for the viscosity growth rate to drop below the first percentage (e.g., 30%) of the initial stage. Based on this actual time, adjust the baseline time to obtain the final set time. For example: extract the viscosity-time curves of the same formulation product from the past 3 months, and mark the average viscosity growth rate V0 in the initial stage (first 5 minutes) of each curve. Calculate the actual time corresponding to when the viscosity growth rate drops to the first percentage (e.g., 30%) of V0 for each curve, and take the average T of these times. 实 Calculate T 实 If the deviation rate from the baseline duration exceeds 15%, a correction value is calculated: Tactual × (1 + deviation coefficient), where the deviation coefficient is dynamically adjusted based on the batch stability of the raw materials (0.05-0.1). This correction value is then used to adjust the baseline duration, for example, by taking the weighted average of the correction value and the baseline duration, to finally obtain the set duration.

[0052] This embodiment takes into account the influence of raw material ratio, equipment parameters, and actual production patterns. It can effectively avoid improper data acquisition timing due to the deviation between theoretical and actual values, improve the effectiveness of multimodal data, provide an accurate time reference for subsequent state identification, and enhance the stability and adaptability of the control method.

[0053] Optionally, the step of dynamically weighting the spatiotemporal correlation features and apparent viscosity temporal features into a gating attention mechanism to obtain a multimodal representation vector includes:

[0054] The spatiotemporal correlation features and the apparent viscosity time series features are dimensionally aligned and mapped to the same dimensional space through linear transformation.

[0055] The dynamic weights of spatiotemporal correlation features and apparent viscosity time-series features are calculated through a gating mechanism, where the gating parameters are adaptively adjusted based on the correlation between features.

[0056] Based on the dynamic weights, the spatiotemporal correlation features and apparent viscosity time-series features are fused, and the contextual correlation of the feature sequence is captured by combining a bidirectional attention mechanism to generate the multimodal representation vector.

[0057] Because the spatiotemporal correlation features (including spatiotemporal information of spectra, liquid surface, and bubbles) and the apparent viscosity time series features (viscosity dynamic change data) have different dimensions, they are first mapped to the same dimensional space through linear transformation.

[0058] Specifically: spatiotemporal correlation features and apparent viscosity time-series characteristics Mapping to a unified dimension through linear transformation :

[0059]

[0060] in, , This is the weight matrix. , This is a bias term.

[0061] Next, the gating mechanism adaptively adjusts the gating parameters based on the real-time correlation between the two types of features (such as the correlation between viscosity changes and liquid level fluctuations), giving higher weight to features that are more discriminative about the current stirring state, thus highlighting key information.

[0062] Specifically: the correlation score between the two types of features is calculated using a bilinear transformation.

[0063]

[0064] in, The correlation matrix, For bias.

[0065] Dynamic weights are generated based on correlation scores:

[0066]

[0067] in, The sigmoid function maps the score to the interval [0,1]. The weights represent the spatiotemporal correlation features. The weights represent the time-series characteristics of apparent viscosity.

[0068] Next, based on the dynamic weight fusion of the two types of features, a bidirectional attention mechanism simultaneously focuses on the historical and future information of the feature sequence (such as the correlation between the current bubble distribution and previous spectral changes), strengthening contextual relevance and ultimately generating a multimodal representation vector that comprehensively reflects the state of the mixed liquid. Specifically:

[0069] Weighted fusion:

[0070] Bidirectional attention mechanism:

[0071] Historical context: Calculate the attention weights for historical features at the current time step.

[0072]

[0073] in, This is the attention weight matrix. .

[0074] Future context (implemented via bidirectional LSTM):

[0075]

[0076] Context fusion:

[0077]

[0078] in, To balance the weights of the forward and backward contexts.

[0079] Multimodal characterization generation

[0080] The final representation vector is

[0081] MLP stands for Multilayer Perceptron. This indicates feature splicing.

[0082] Optionally, the step of inputting the multimodal representation vector into an improved Transformer temporal classifier and dynamically outputting the stirring state determination result based on a progressive classification threshold includes:

[0083] The improved Transformer time series classifier encodes the relative positions of the input multimodal representation vectors to obtain the time series.

[0084] The window length is adjusted based on the feature fluctuation amplitude of the multimodal representation vector. The time series is divided into multiple windows corresponding to the adjusted window length. Query-key attention is calculated within the window, and gradient flow is maintained through residual connection and layer normalization.

[0085] The feature sequence processed by the multi-layer Transformer encoder is pooled to obtain the global feature vector, which is then input into the fully connected layer and the classification probability distribution of the stirring state is calculated using the softmax activation function.

[0086] When the classification probability of the stirring state output by the time classifier exceeds the progressive classification threshold of the corresponding stage, the output determines that a viscous stable state has been reached; where the progressive classification threshold is calculated based on the current time step and the critical time.

[0087] To address the issue of Transformer's insensitivity to sequence order and enable the model to capture temporal dependencies during the mixing process, this embodiment injects relative positional information into the temporal features. Specifically:

[0088] First, the multimodal representation vector is positionally encoded using sine and cosine functions, as shown in the formula:

[0089]

[0090] in, , For feature dimension, For time steps.

[0091] Then, the window length L is dynamically adjusted according to the fluctuation amplitude of the multimodal representation vector (the larger the fluctuation, the smaller L), so as to achieve fine perception of abrupt features (such as a sudden increase in viscosity).

[0092] The attention weights are calculated within each window of length L, using the following formula:

[0093]

[0094] Among them, the mask matrix By limiting the attention scope to within the window, the ability to capture local temporal patterns is enhanced.

[0095] In addition, residual connections are also used ( Layer normalization (LayerNorm) prevents gradient vanishing and ensures the stability of deep network training.

[0096] Next, pooling operations (such as average pooling or CLS labeling) compress the sequence features output from the multi-layer Transformer into a fixed-dimensional global feature vector. This global feature vector is then input into a fully connected layer and transformed into a probability distribution of the stirring state using a softmax function. ,in This indicates different state categories (such as "unstable", "critically stable", "densely stable").

[0097] The formula for calculating the progressive classification threshold is:

[0098]

[0099] in, For the current time step, This is a preset critical time, which is the boundary between the viscous steady state and the non-viscous steady state. Control the threshold growth rate, Adjust the threshold curve offset.

[0100] As stirring continues, the progressive classification threshold increases over time. The judgment result is only output when the model's predicted probability of the "dense and stable" state exceeds the current threshold, thus avoiding misjudgments caused by early fluctuations.

[0101] In this embodiment, by employing relative position encoding and a dynamic window mechanism, the model's ability to capture time-dependent relationships and local abrupt changes during the stirring process is significantly improved. Simultaneously, the progressive threshold mechanism adapts to the dynamic characteristics of the phase change material stirring process, reducing the risk of misjudgment, and making decisions more reliable, especially near the critical point.

[0102] Optionally, the adjustment window length based on the feature fluctuation amplitude of the multimodal representation vector includes:

[0103] The mean distance between the feature vectors of adjacent time steps is calculated by using a sliding window to obtain the fluctuation coefficient, which is then used as the feature fluctuation amplitude; and the Euclidean distance between the actual raw material ratio and the standard ratio of the mixed liquid is calculated.

[0104] If the fluctuation coefficient is greater than the preset fluctuation threshold and the Euclidean distance is less than or equal to the preset deviation threshold, then the window length is shortened to a second percentage of the initial value.

[0105] If the fluctuation coefficient is less than or equal to a preset fluctuation threshold and the Euclidean distance is greater than a preset deviation threshold, then the window length is shortened proportionally according to the deviation rate; wherein the deviation rate is calculated based on the Euclidean distance.

[0106] If the fluctuation coefficient is greater than the preset fluctuation threshold and the Euclidean distance is greater than the preset deviation threshold, then the shorter window length of the two adjustment methods mentioned above shall be selected.

[0107] If the fluctuation coefficient is less than or equal to a preset fluctuation threshold and the Euclidean distance is less than or equal to a preset deviation threshold, then the window length is set to remain at its initial value.

[0108] The mean Euclidean distance (e.g., Euclidean distance) between multimodal representation vectors at adjacent time steps is calculated using the sliding window method to obtain the volatility coefficient (i.e., a quantitative indicator of the characteristic volatility amplitude). The formula is:

[0109]

[0110] in, The length of the sliding window. Let t be the multimodal representation vector at time t. The Euclidean distance intuitively reflects the degree of drastic change of features over time.

[0111] Meanwhile, the raw material ratio is a core factor determining the reaction characteristics of the mixed liquid. Deviations between the actual and standard ratios can lead to shifts in key features such as reaction rate and viscosity variation. For example, when the proportion of polyurethane solution is too high, its crosslinking reaction with paraffin is delayed, and the temporal pattern of the multimodal characterization vector will differ significantly from the characteristics under the standard ratio. When the raw material ratio deviation is large, the degree of variation in the feature sequence increases. If a fixed window length is still used, it may not be able to accurately capture local feature mutations caused by abnormal ratios. Therefore, in this embodiment, the window length is shortened proportionally to the deviation rate, which can improve the sensitivity to subtle feature changes under abnormal ratios and ensure that the attention mechanism can focus on the key patterns caused by ratio fluctuations.

[0112] In addition, there may be a coupling relationship between raw material ratio deviation and characteristic fluctuation amplitude (such as ratio deviation causing characteristic fluctuation). Combining the two to adjust the window length can avoid the limitations of adjusting a single parameter, so that the window division can simultaneously adapt to the stability of raw material input and real-time characteristic changes, thereby enhancing the adaptability of the time series classifier to complex working conditions.

[0113] Specifically: The deviation between the actual raw material ratio and the standard ratio is quantified by Euclidean distance, using the following formula:

[0114]

[0115] in, For standard matching vectors, This is the actual matching vector. The larger the value, the more significant the deviation of the raw material ratio from the standard.

[0116] Deviation rate ( (This is a preset deviation threshold).

[0117] when and At this time, the window length L is shortened to a second percentage of the initial value (such as 60%-80%) to enhance the ability to locally capture high-frequency fluctuation characteristics.

[0118] when and At that time, the window length L is shortened proportionally according to the deviation rate (e.g., L is shortened by 5% for every 10% increase in the deviation rate) to adapt to the characteristic distribution shift caused by abnormal raw material ratio.

[0119] when and At that time, the shorter window length is selected from the two methods of shortening by the second percentage and shortening by the same period of deviation rate, so as to give priority to the dual sensitivity to capture drastic fluctuations and raw material anomalies.

[0120] when and At the same time, the window length remains at its initial value to capture the temporal characteristics in a stable state from a global perspective, thus avoiding waste of computing resources.

[0121] This embodiment adjusts the window length by combining the amplitude of feature fluctuations and the deviation of raw material ratios. This responds to real-time feature changes while also taking into account the stability of raw material inputs, making the window division more closely match the complex working conditions of actual production. Furthermore, through a tiered adjustment strategy, the window is shortened to improve the accuracy of local feature resolution when there are feature fluctuations or raw material anomalies, while the window length is maintained in a stable state to preserve computational efficiency.

[0122] like Figure 2 As shown, this embodiment of the invention also discloses a production control system for phase change materials, the system including an acquisition unit 1001, a feature extraction unit 1002, and a stability determination unit 1003;

[0123] The acquisition unit 1001 acquires multimodal data of the mixed liquid during the stirring process when the stirring time of the mixed liquid reaches the set time, including reflectance spectral characteristics, liquid surface fluctuation amplitude, bubble distribution characteristics, and torque feedback data.

[0124] The feature extraction unit 1002 uses a spatiotemporal convolutional neural network to extract spatiotemporal correlation features from reflectance spectrum features, liquid surface fluctuation amplitude, and bubble distribution features. It also uses a rheological feature extraction module to extract apparent viscosity time-series features from torque feedback data. The spatiotemporal correlation features and apparent viscosity time-series features are input into a gating attention mechanism for dynamic weighting to obtain a multimodal representation vector.

[0125] The stability determination unit 1003 inputs the multimodal representation vector into an improved Transformer temporal classifier and dynamically outputs the stirring state determination result based on a progressive classification threshold. When the determination reaches a viscous stable state, a process switching signal is triggered. The temporal classifier introduces relative position encoding to enhance the correlation of local features.

[0126] Optionally, the set duration is determined in the following manner:

[0127] The first theoretical time required for the reaction system to reach a uniform mixing state is calculated based on the raw material ratio data of the mixed liquid. The second theoretical time required for the fluid to reach a uniform mixing state is calculated by combining the geometric parameters of the stirrer and the fluid dynamics formula. The baseline time is calculated by weighting the first theoretical time and the second theoretical time.

[0128] By analyzing the viscosity change curve in historical production data, the duration during which the viscosity growth rate drops below the first percentage of the initial stage is determined. Based on this duration, the baseline duration is corrected to obtain the set duration.

[0129] Optionally, the acquisition unit 1001 specifically:

[0130] The spatiotemporal correlation features and the apparent viscosity time series features are dimensionally aligned and mapped to the same dimensional space through linear transformation.

[0131] The dynamic weights of spatiotemporal correlation features and apparent viscosity time-series features are calculated through a gating mechanism, where the gating parameters are adaptively adjusted based on the correlation between features.

[0132] Based on the dynamic weights, the spatiotemporal correlation features and apparent viscosity time-series features are fused, and the contextual correlation of the feature sequence is captured by combining a bidirectional attention mechanism to generate the multimodal representation vector.

[0133] Optionally, the stability determination unit 1003 specifically:

[0134] The improved Transformer time series classifier encodes the relative positions of the input multimodal representation vectors to obtain the time series.

[0135] The window length is adjusted based on the feature fluctuation amplitude of the multimodal representation vector. The time series is divided into multiple windows corresponding to the adjusted window length. Query-key attention is calculated within the window, and gradient flow is maintained through residual connection and layer normalization.

[0136] The feature sequence processed by the multi-layer Transformer encoder is pooled to obtain the global feature vector, which is then input into the fully connected layer and the classification probability distribution of the stirring state is calculated using the softmax activation function.

[0137] When the classification probability of the stirring state output by the time classifier exceeds the progressive classification threshold of the corresponding stage, the output determines that a viscous stable state has been reached; where the progressive classification threshold is calculated based on the current time step and the critical time.

[0138] Optionally, the stability determination unit 1003 specifically:

[0139] The mean distance between the feature vectors of adjacent time steps is calculated by using a sliding window to obtain the fluctuation coefficient, which is then used as the feature fluctuation amplitude; and the Euclidean distance between the actual raw material ratio and the standard ratio of the mixed liquid is calculated.

[0140] If the fluctuation coefficient is greater than the preset fluctuation threshold and the Euclidean distance is less than or equal to the preset deviation threshold, then the window length is shortened to a second percentage of the initial value.

[0141] If the fluctuation coefficient is less than or equal to a preset fluctuation threshold and the Euclidean distance is greater than a preset deviation threshold, then the window length is shortened proportionally according to the deviation rate; wherein the deviation rate is calculated based on the Euclidean distance.

[0142] If the fluctuation coefficient is greater than the preset fluctuation threshold and the Euclidean distance is greater than the preset deviation threshold, then the shorter window length of the two adjustment methods mentioned above shall be selected.

[0143] If the fluctuation coefficient is less than or equal to a preset fluctuation threshold and the Euclidean distance is less than or equal to a preset deviation threshold, then the window length is set to remain at its initial value.

[0144] This invention also discloses an electronic device, comprising: a memory storing executable program code; a processor coupled to the memory; the processor calling the executable program code stored in the memory to execute the method described in Embodiment 1.

[0145] This invention also discloses a computer storage medium storing a computer program, which is executed by a processor as described in Embodiment 1.

[0146] This invention also discloses a computer program product comprising a computer program for performing the method as described in any of the preceding claims.

[0147] The above description is merely a preferred embodiment and the technical principles employed in this application. Those skilled in the art will understand that this application is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions can be made without departing from the scope of protection of this application. Therefore, although this application has been described in detail through the above embodiments, this application is not limited to the above embodiments, and may include many other equivalent embodiments without departing from the concept of this application, the scope of which is determined by the scope of the appended claims.

Claims

1. A method for controlling the production of phase change materials, characterized in that, Includes the following steps: When the stirring time of the mixed liquid reaches the set time, multimodal data of the mixed liquid during the stirring process are acquired, including reflectance spectral characteristics, liquid surface fluctuation amplitude, bubble distribution characteristics, and torque feedback data. Spatiotemporal convolutional neural networks are used to extract spatiotemporal correlation features from reflectance spectrum features, liquid surface fluctuation amplitude, and bubble distribution features. A rheological feature extraction module is used to extract apparent viscosity time-series features from torque feedback data. The spatiotemporal correlation features and apparent viscosity time-series features are input into a gating attention mechanism for dynamic weighting to obtain a multimodal representation vector. The multimodal representation vector is input into an improved Transformer temporal classifier, which dynamically outputs the stirring state determination result based on a progressive classification threshold. When the determination reaches a viscous stable state, a process switching signal is triggered. The temporal classifier introduces relative position encoding to enhance the correlation of local features. The spatiotemporal correlation features and apparent viscosity temporal features are input into a gating attention mechanism for dynamic weighting to obtain a multimodal representation vector, including: The spatiotemporal correlation features and the apparent viscosity time series features are dimensionally aligned and mapped to the same dimensional space through linear transformation. The dynamic weights of spatiotemporal correlation features and apparent viscosity time-series features are calculated through a gating mechanism, where the gating parameters are adaptively adjusted based on the correlation between features. Based on the dynamic weights, the spatiotemporal correlation features and apparent viscosity time-series features are fused, and the contextual correlation of the feature sequence is captured by combining a bidirectional attention mechanism to generate the multimodal representation vector.

2. The production control method for a phase change material according to claim 1, characterized in that: The set duration is determined in the following manner: The first theoretical time required for the reaction system to reach a uniform mixing state is calculated based on the raw material ratio data of the mixed liquid. The second theoretical time required for the fluid to reach a uniform mixing state is calculated by combining the geometric parameters of the stirrer and the fluid dynamics formula. The baseline time is calculated by weighting the first theoretical time and the second theoretical time. By analyzing the viscosity change curve in historical production data, the duration during which the viscosity growth rate drops below the first percentage of the initial stage is determined. Based on this duration, the baseline duration is corrected to obtain the set duration.

3. The production control method for a phase change material according to claim 1, characterized in that: The multimodal representation vector is input into an improved Transformer temporal classifier, and the stirring state determination result is dynamically output based on a progressive classification threshold, including: The improved Transformer time series classifier encodes the relative positions of the input multimodal representation vectors to obtain the time series. The window length is adjusted based on the feature fluctuation amplitude of the multimodal representation vector. The time series is divided into multiple windows corresponding to the adjusted window length. Query-key attention is calculated within the window, and gradient flow is maintained through residual connection and layer normalization. The feature sequence processed by the multi-layer Transformer encoder is pooled to obtain the global feature vector, which is then input into the fully connected layer and the classification probability distribution of the stirring state is calculated using the softmax activation function. When the classification probability of the stirring state output by the time classifier exceeds the progressive classification threshold of the corresponding stage, the output determines that a viscous stable state has been reached; where the progressive classification threshold is calculated based on the current time step and the critical time.

4. The production control method for a phase change material according to claim 3, characterized in that: Adjusting the window length based on the feature fluctuation amplitude of multimodal representation vectors includes: The mean distance between the feature vectors of adjacent time steps is calculated by using a sliding window to obtain the fluctuation coefficient, which is then used as the feature fluctuation amplitude; and the Euclidean distance between the actual raw material ratio and the standard ratio of the mixed liquid is calculated. If the fluctuation coefficient is greater than a preset fluctuation threshold and the Euclidean distance is less than or equal to a preset deviation threshold, then the window length is shortened to a second percentage of the initial value. If the fluctuation coefficient is less than or equal to a preset fluctuation threshold and the Euclidean distance is greater than a preset deviation threshold, then the window length is shortened proportionally according to the deviation rate; wherein the deviation rate is calculated based on the Euclidean distance. If the fluctuation coefficient is greater than the preset fluctuation threshold and the Euclidean distance is greater than the preset deviation threshold, then the shorter window length of the two adjustment methods mentioned above shall be selected. If the fluctuation coefficient is less than or equal to a preset fluctuation threshold and the Euclidean distance is less than or equal to a preset deviation threshold, then the window length is set to remain at its initial value.

5. A production control system for phase change materials, characterized in that, The system includes an acquisition unit, a feature extraction unit, and a stability determination unit; The acquisition unit acquires multimodal data of the mixed liquid during the stirring process when the stirring time of the mixed liquid reaches the set time, including reflectance spectral characteristics, liquid surface fluctuation amplitude, bubble distribution characteristics, and torque feedback data. The feature extraction unit uses a spatiotemporal convolutional neural network to extract spatiotemporal correlation features from reflectance spectrum features, liquid surface fluctuation amplitude, and bubble distribution features. It also uses a rheological feature extraction module to extract apparent viscosity time-series features from torque feedback data. The spatiotemporal correlation features and apparent viscosity time-series features are input into a gating attention mechanism for dynamic weighting to obtain a multimodal representation vector. The stability determination unit inputs the multimodal representation vector into an improved Transformer temporal classifier and dynamically outputs the stirring state determination result based on a progressive classification threshold. When the determination reaches a viscous stable state, a process switching signal is triggered. The temporal classifier introduces relative position encoding to enhance the correlation of local features. The acquisition unit specifically includes: The spatiotemporal correlation features and the apparent viscosity time series features are dimensionally aligned and mapped to the same dimensional space through linear transformation. The dynamic weights of spatiotemporal correlation features and apparent viscosity time-series features are calculated through a gating mechanism, where the gating parameters are adaptively adjusted based on the correlation between features. Based on the dynamic weights, the spatiotemporal correlation features and apparent viscosity time-series features are fused, and the contextual correlation of the feature sequence is captured by combining a bidirectional attention mechanism to generate the multimodal representation vector.

6. The production control system for phase change materials according to claim 5, characterized in that: The set duration is determined in the following manner: The first theoretical time required for the reaction system to reach a preliminary mixing state is calculated based on the raw material ratio data of the mixed liquid. The second theoretical time required for the fluid to reach uniform mixing is calculated by combining the geometric parameters of the stirrer and the fluid dynamics formula. The baseline time is obtained by weighted calculation based on the first theoretical time and the second theoretical time. By analyzing the viscosity change curve in historical production data, the duration during which the viscosity growth rate drops below the first percentage of the initial stage is determined. Based on this duration, the baseline duration is corrected to obtain the set duration.

7. An electronic device, comprising: Memory containing executable program code; A processor coupled to the memory; characterized in that: the processor calls the executable program code stored in the memory to perform the method as described in any one of claims 1-4.

8. A computer storage medium storing a computer program, characterized in that: The computer program is executed by the processor to perform the method as described in any one of claims 1-4.

9. A computer program product, characterized in that: The computer program product includes a computer program for performing the method as described in any one of claims 1 to 4.

Images