Paper cup hot-press forming temperature self-adaptive control system based on deep learning model

Abstract

The application discloses a paper cup hot-press forming temperature self-adaptive control system based on a deep learning model, relates to the field of temperature self-adaptive control, and comprises a record definition unit, a graph construction unit and a graph training unit, is used for traversing P high-quality production records, obtaining P space-time graph structures, taking the P space-time graph structures as training samples of a time series graph neural model, iteratively training a temperature distribution prediction model, and a self-adaptive control unit, which is used for performing self-adaptive control of the paper cup hot-press forming temperature in real-time production operation based on the temperature distribution prediction model. The application can continuously guide the temperature trajectory to approach an optimal forming path under the influence of uncertain factors such as material property changes, environmental disturbances or equipment state drift, effectively suppresses local overheating and underheating phenomena, and improves the wall thickness consistency, edge bonding strength and overall appearance yield of the paper cup.

Description

Adaptive control system for paper cup thermoforming temperature based on deep learning model Technical Field

[0001] This invention relates to the field of temperature adaptive control, specifically to a temperature adaptive control system for paper cup hot pressing based on a deep learning model. Background Technology

[0002] As disposable environmentally friendly beverage containers, paper cups' structural strength, sealing performance, and appearance quality are directly affected by the thermoforming process during production. During thermoforming, the temperature distribution in different areas of the mold plays a decisive role in fiber plasticization, adhesive activation, and cooling and setting. Traditional temperature control methods often employ zoned constant temperature control or preset temperature curves, relying on manual experience to set heating parameters. This makes it difficult to cope with dynamic disturbances such as batch differences in raw materials, fluctuations in ambient temperature and humidity, and mold wear. This can easily lead to localized overheating causing paper base carbonization, or insufficient temperature resulting in weak adhesion, edge curling, and cracking.

[0003] The prior art disclosed in patent publication number CN103057102B is a hot press head and hot pressing device that uses closed-loop temperature feedback control to stabilize the temperature of multiple hot press heads. Furthermore, some devices in the prior art introduce multi-point temperature feedback and PID closed-loop control to achieve tracking and adjustment of the set temperature. However, their control target is still a static or segmented fixed temperature value, failing to consider the differentiated requirements of different process stages (such as stretching, pressing, and curling) for the dynamic temperature change path. In addition, the mold surface is a non-planar structure, especially in the transition area between the cup body and the bottom, where heat conduction is significantly affected by geometric curvature. Traditional uniform temperature control strategies based on spatial distance cannot accurately reflect the actual heat field distribution, leading to unstable molding quality in complex curved areas.

[0004] In recent years, machine learning models have been used to predict temperature trends and assist in adjusting heating parameters. However, most methods treat each temperature measurement point as an independent variable, lacking modeling of the spatial topological relationship of the mold surface and failing to reflect the conduction coupling effect of heat between adjacent hot-pressing positions. At the same time, model training often relies on fixed thresholds to select samples, ignoring the impact of process drift on the definition of "high quality" during production. This results in a lack of dynamic adaptation to temperature control in paper cup hot-pressing equipment, making it difficult to maintain the optimal temperature distribution under changing operating conditions. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an adaptive temperature control system for paper cup hot pressing based on a deep learning model, which solves the technical problems mentioned in the background by introducing spatiotemporal graph structure modeling.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] The paper cup hot-pressing temperature adaptive control system based on a deep learning model includes the following application units:

[0008] Record definition unit, used to define paper cup production records;

[0009] The graph construction unit is used to construct a spatiotemporal graph structure of the paper cup thermoforming process based on the paper cup production records.

[0010] The graph training unit is used to traverse P high-quality production records, repeatedly execute the construction of spatiotemporal graph structures until P spatiotemporal graph structures are obtained, and use the P spatiotemporal graph structures as training samples for the time series graph neural model to iteratively train the temperature distribution prediction model.

[0011] The training objective of the temperature distribution prediction model is defined as: given the non-temperature parameter time series of H hot-pressing locations, predict the recommended temperature evolution sequence of each hot-pressing location within a future time window;

[0012] An adaptive control unit is used to adaptively control the temperature of paper cup hot pressing during real-time production operations based on a temperature distribution prediction model.

[0013] In some embodiments, the graph building unit includes:

[0014] The threshold calculation subunit is used to select N paper cup production records within a production cycle from the storage database of paper cup production records, and calculate the dynamic quality threshold within the production cycle.

[0015] The cycle selection sub-unit is used to sequentially traverse the M production cycles in the storage database based on a preset time window and sliding step size, and select N×M paper cup production records of the M production cycles.

[0016] The high-quality screening subunit is used to screen out P high-quality production records from N×M paper cup production records in M ​​production cycles based on a dynamic quality threshold.

[0017] The sequence construction subunit is used to construct the non-temperature parameter time series and temperature parameter time series of H hot-pressing locations for each high-quality production record;

[0018] The spatiotemporal graph generation subunit is used to perform graph structure construction on the non-temperature parameter time series and temperature parameter time series of H hot-pressing locations, and generate a spatiotemporal graph structure composed of H hot-pressing locations.

[0019] In some embodiments, the calculation steps of the threshold calculation subunit are as follows:

[0020] B1-1. Based on N paper cup production records within the production cycle, anchor the finished product quality parameters of each paper cup production record;

[0021] B1-2. Based on the finished product quality parameters of each paper cup production record, calculate the quality score of N paper cup production records;

[0022] B1-3. Calculate the dynamic quality threshold within the production cycle based on the quality scores of N paper cup production records.

[0023] In some embodiments, calculating the dynamic quality threshold within the production cycle includes:

[0024] B1-3-1. Sort the quality scores of N paper cup production records in ascending order to obtain an ordered quality score sequence.

[0025] B1-3-2. Based on the preset proportion of high-quality samples, anchor the target quantile in the ordered quality scoring sequence;

[0026] B1-3-3, Extract the quality score corresponding to the target quantile as the dynamic quality threshold within the production cycle.

[0027] In some embodiments, the screening steps of the high-quality screening subunit are as follows:

[0028] B3-1. Based on N paper cup production records for each production cycle, determine M dynamic quality thresholds;

[0029] B3-2. Perform K-value clustering on the M dynamic quality thresholds to obtain K threshold clusters; where each threshold cluster contains several dynamic quality thresholds.

[0030] B3-3. Calculate the threshold variance of each threshold cluster based on the several dynamic quality thresholds contained in each threshold cluster.

[0031] B3-4. Among the K threshold clusters, select the threshold cluster corresponding to the largest threshold variance and denote it as the target threshold cluster.

[0032] B3-5. Calculate the average threshold of all dynamic quality thresholds in the target threshold cluster, and denote it as the target threshold.

[0033] B3-6. From N×M paper cup production records across M production cycles, select P paper cup production records whose quality scores exceed the target threshold and define them as high-quality production records.

[0034] In some embodiments, M dynamic quality thresholds are clustered using K-values ​​to obtain K threshold clusters, including:

[0035] B3-2-1. From M dynamic quality thresholds, randomly select K dynamic quality thresholds as the first round cluster centers;

[0036] B3-2-2. For each dynamic quality threshold, calculate its threshold deviation from the K first-round cluster centers and assign it to the first-round cluster center with the smallest threshold deviation.

[0037] B3-2-3. Traverse all dynamic quality thresholds until each dynamic quality threshold is assigned to the first-round cluster center with the smallest threshold deviation, forming K initial threshold clusters.

[0038] B3-2-4. Calculate the average threshold of each cluster based on all dynamic quality thresholds within each initial threshold cluster, and anchor the average threshold as the cluster center for the next round.

[0039] B3-2-5. Calculate the threshold deviation between the cluster centers of the next round and the cluster centers of the first round;

[0040] B3-2-6. If the threshold deviation is greater than the preset threshold, then all dynamic quality thresholds will be iteratively allocated based on the next round of cluster centers.

[0041] B3-2-7. Repeat the iterative allocation until all threshold deviations are less than the preset threshold, to obtain K threshold clusters;

[0042] B3-2-8. Perform the iterative allocation on each of the M dynamic quality thresholds to obtain K threshold clusters.

[0043] In some embodiments, the construction steps of the sequence construction subunit are as follows:

[0044] B4-1. Select the target hot-pressing location from among the hot-pressing locations in the high-quality production record;

[0045] B4-2. Obtain the non-temperature process parameters of the target hot-pressing position at T production timestamps; where T production timestamps represent consecutive timestamps on the time axis;

[0046] B4-3. Characterize the non-temperature process parameters into a non-temperature process vector with T production timestamps;

[0047] B4-4. Arrange the T non-temperature process vectors according to the time sequence of the T production timestamps to construct the time series of non-temperature parameters for the target hot pressing position.

[0048] B4-5. Obtain the temperature control parameters of the target hot-pressing location at T production timestamps;

[0049] B4-6. Characterize the temperature control parameters into temperature features of T production timestamps;

[0050] B4-7. Arrange the T temperature features according to the time sequence of the T production timestamps to construct a time series of temperature parameters for the target hot pressing location;

[0051] B4-8. Traverse the high-quality production records and repeatedly execute sequence construction until you obtain the non-temperature parameter time series and temperature parameter time series of H hot pressing positions.

[0052] Both the non-temperature parameter time series and the temperature parameter time series are labeled with the three-dimensional coordinates of their hot-pressing locations.

[0053] In some embodiments, the construction steps of the spatiotemporal graph generation subunit are as follows:

[0054] B5-1. Define the three-dimensional coordinates of the hot-pressing position as nodes, define the non-temperature parameter time series as node features, and define the temperature parameter time series as node labels, thus forming H basic nodes.

[0055] B5-2. Based on the three-dimensional coordinates of the bottom of the paper cup, divide the H basic nodes into cup bottom nodes and cup body nodes.

[0056] B5-3. In the cup bottom nodes, calculate the Euclidean distance between any two cup bottom nodes based on the three-dimensional coordinates of the cup bottom nodes;

[0057] B5-4. If the Euclidean distance is less than the set threshold, a connecting edge is established between the two cup bottom nodes, and the Euclidean distance is inversely proportional to the first heat conduction weight of the connecting edge to obtain the cup bottom diagram structure.

[0058] B5-5. In the cup body nodes, based on the three-dimensional coordinates of the cup body nodes and the cup body curvature corresponding to the three-dimensional coordinates, calculate the weighted geodesic distance between any two cup body nodes.

[0059] B5-6. If the weighted geodesic distance is less than the set threshold, a directed edge is established between the two cup body nodes, and the geodesic distance is inversely proportional to the second heat conduction weight of the connecting edge to obtain the cup body diagram structure.

[0060] B5-7. Traverse the H hot-pressing positions in the high-quality production record, merge the cup bottom diagram structure and the cup body diagram structure to generate the spatiotemporal diagram structure of the H hot-pressing positions within the paper cup hot-pressing cycle.

[0061] In some embodiments, the control steps of the adaptive control unit are as follows:

[0062] C1. Real-time acquisition of non-temperature process parameters for each hot-pressing position within a standard time period;

[0063] C2. Construct a real-time spatiotemporal diagram structure from the non-temperature process parameters of each hot pressing position over a standard duration;

[0064] C3. Input the real-time spatiotemporal graph structure into the temperature distribution prediction model and output the recommended temperature evolution sequence of each hot-pressed region within the future time window;

[0065] C4. Based on the recommended temperature evolution sequence of each hot-pressing region within a future time window, adaptive control of the hot-pressing temperature of paper cups is performed.

[0066] This invention provides an adaptive control system for the thermoforming temperature of paper cups based on a deep learning model, which has the following beneficial effects:

[0067] This invention achieves spatiotemporal adaptive adjustment of the paper cup hot-pressing temperature control by constructing a spatiotemporal graph structure that integrates the three-dimensional coordinates of the hot-pressing position and the production timestamp, and by introducing an evaluation mechanism based on a dynamic quality threshold. Compared with traditional zoned constant temperature or fixed curve control methods, this system can generate recommended temperature sequences based on real-time process conditions, significantly improving the uniformity of temperature distribution in complex curved areas.

[0068] Furthermore, by discretizing the hot-pressing position nodes on the mold surface and establishing undirected and directed connections with the bottom plane and the curved surface of the cup body respectively, the heat conduction path on the non-uniform mold surface is effectively expressed. This overcomes the problem of treating the temperature measurement point as an isolated unit in the traditional method, making the temperature prediction closer to the actual heat conduction behavior.

[0069] Furthermore, by dynamically calculating quality thresholds according to the production cycle and clustering to identify significant process fluctuation cycles, condition-adaptive screening of high-quality production records was achieved. This enabled the model training to not only reflect static quality standards but also demonstrate the system's anti-interference capability under disturbance conditions, thereby enhancing the stability of the model output in the real production line environment.

[0070] Ultimately, by using the recommended temperature evolution sequence as a dynamic target input to the underlying controller and combining it with actual temperature feedback to form a closed-loop regulation, the temperature trajectory can be continuously guided to approach the optimal forming path under the influence of uncertain factors such as changes in material properties, environmental disturbances, or equipment state drift. This effectively suppresses local overheating and underheating, and improves the consistency of paper cup wall thickness, edge bonding strength, and overall appearance yield. Attached Figure Description

[0071] Figure 1 is a structural block diagram of the paper cup hot pressing temperature adaptive control system based on a deep learning model according to the present invention.

[0072] Figure 2 is a schematic diagram of the control flow of the paper cup hot pressing temperature adaptive control system based on a deep learning model according to the present invention.

[0073] Figure 3 is a schematic diagram of the sub-unit structure of the graph construction unit of the present invention;

[0074] Figure 4 is a schematic diagram of the generation process of an embodiment of the spatiotemporal graph structure described in this invention;

[0075] Figure 5 is a schematic diagram of the extraction process of the dynamic quality threshold described in this invention;

[0076] Figure 6 is a schematic diagram of the definition process of the high-quality production record according to the present invention;

[0077] Figure 7 is a schematic diagram of the construction process of another embodiment of the spatiotemporal graph structure described in this invention. Detailed Implementation

[0078] 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 embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0079] Example 1: Referring to Figures 1 and 2, the present invention provides an adaptive control system for paper cup hot pressing temperature based on a deep learning model, including the following application units:

[0080] Record definition unit, used to define paper cup production records;

[0081] The paper cup production record includes: non-temperature process parameters, temperature control parameters, and finished product quality parameters;

[0082] The non-temperature process parameters include: the three-dimensional coordinates, pressure, humidity, raw material characteristics, and molding time of each hot-pressing position of the paper cup mold;

[0083] The temperature control parameters include: the three-dimensional coordinates and heating temperature of each hot-pressing position of the paper cup mold;

[0084] The finished product quality parameters include: the thickness, strength, and external dimensions of the paper cup area corresponding to each hot-pressing position of the paper cup mold;

[0085] Specifically, in this embodiment, the paper cup production record represents a set of process data collected by a distributed sensor system within a complete paper cup thermoforming cycle. Each production record corresponds to an independently produced paper cup unit, and its data structure uses timestamps as indexes to record the real-time process parameters of each thermoforming position on the mold surface and the quality inspection results of the final product during each process stage, such as mold closing, heating and pressurizing, and cooling and demolding.

[0086] The hot-pressing position refers to the physical location on the surface of the paper cup forming mold. In this embodiment, based on the three-dimensional CAD model of the mold, the mold surface is meshed to generate a high-density set of discrete spatial nodes, which are used to characterize the temperature and pressure distribution characteristics of the mold surface. The spatial nodes cover key forming areas such as the bottom, wall, and rim of the cup, achieving spatial coverage of the working surface of the mold.

[0087] The three-dimensional coordinates of all hot-pressing locations are precisely calibrated using a CAD model and used to construct the spatial topology between nodes. Non-temperature process parameters (including pressure, humidity, and raw material characteristics) and temperature control parameters are synchronously acquired at a sampling frequency of 10Hz to ensure time alignment of the timing data. Finished product quality parameters are measured point-by-point in the forming area corresponding to each discretized spatial node using a machine vision system and a laser thickness gauge after the product is completed, obtaining the thickness, strength, and dimensional deviation of the paper cup at that location.

[0088] The graph construction unit is used to construct a spatiotemporal graph structure of the paper cup thermoforming process based on the paper cup production records.

[0089] Specifically, the graph unit includes the following sub-units:

[0090] The threshold calculation subunit is used to select N paper cup production records within a production cycle from the storage database of paper cup production records, and calculate the dynamic quality threshold within the production cycle.

[0091] The cycle selection sub-unit is used to sequentially traverse the M production cycles in the storage database based on a preset time window and sliding step size, and select N×M paper cup production records of the M production cycles.

[0092] The high-quality screening subunit is used to screen out P high-quality production records from N×M paper cup production records in M ​​production cycles based on a dynamic quality threshold.

[0093] The sequence construction subunit is used to construct the non-temperature parameter time series and temperature parameter time series of H hot-pressing locations for each high-quality production record;

[0094] The spatiotemporal graph generation subunit is used to perform graph structure construction on the non-temperature parameter time series and temperature parameter time series of H hot-pressing locations, and generate a spatiotemporal graph structure composed of H hot-pressing locations;

[0095] The graph training unit is used to traverse P high-quality production records, repeatedly execute the construction of spatiotemporal graph structures until P spatiotemporal graph structures are obtained, and use the P spatiotemporal graph structures as training samples for the time series graph neural model to iteratively train the temperature distribution prediction model.

[0096] The training objective of the temperature distribution prediction model is defined as: given the non-temperature parameter time series of H hot-pressing locations, predict the recommended temperature evolution sequence of each hot-pressing location within a future time window;

[0097] It should be noted that the temporal graph neural model is a deep learning architecture that jointly models spatial topology and temporal evolution, suitable for spatiotemporal prediction tasks. In this embodiment, a hybrid model structure combining a graph convolutional network (GCN) and a gated recurrent unit (GRU), namely a spatial-temporal graph neural network (ST-GNN), is used to learn the spatial and temporal evolution of each hot-pressing position during the paper cup hot-pressing process. Specifically, the GCN extracts spatial features, while the GRU acts as a temporal encoder to capture the temporal dynamics of process parameters.

[0098] Its iterative training process is as follows:

[0099] A1. Data preparation: P spatiotemporal graph structures constructed from P high-quality production records are used as the training sample set. Each sample contains input (non-temperature parameter time series) and label (actual temperature parameter time series).

[0100] A2. Model forward propagation: Input each spatiotemporal graph structure into the ST-GNN model, and the model outputs the predicted recommended temperature evolution sequence;

[0101] A3. Loss Calculation: The mean squared error (MSE) loss function is used to calculate the error between the recommended temperature evolution sequence and the actual temperature parameter time series.

[0102] ;

[0103] in:

[0104] This represents the error between the recommended temperature evolution sequence and the actual temperature parameter time series;

[0105] P represents the number of high-quality production records, and p represents the index of high-quality production records;

[0106] H represents the number of hot-pressing positions on the surface of the paper cup mold, and h represents the index of the hot-pressing position;

[0107] F represents the length of the future time window (i.e., the number of prediction steps), and f represents the index of the future time step;

[0108] This represents the recommended heating temperature predicted by the model for the p-th record, the h-th hot pressing location, and the f-th future time step.

[0109] This represents the actual heating temperature at the p-th record, the h-th hot pressing position, and the f-th future time step.

[0110] A4. Backpropagation and parameter update: Minimize the loss function using a gradient descent algorithm (such as the Adam optimizer) to update the model parameters;

[0111] A5. Iterative convergence judgment: When the validation set loss no longer decreases for several consecutive rounds, or when the preset maximum number of training rounds is reached, training is stopped, and the final temperature distribution prediction model is obtained.

[0112] The future time window refers to the time range of several future time steps predicted by the model from the current moment; in this embodiment, F=5.

[0113] The recommended temperature evolution sequence represents the heating temperature evolution sequence that the hot-pressing positions should follow within a future time window to achieve the goal of high-quality molding, as predicted by the model. This sequence is not a single target value, but a time-varying sequence of control commands, reflecting the optimal evolution path of temperature under dynamic process conditions.

[0114] Furthermore, this recommended temperature evolution sequence is trained based on historical high-quality production data, reflecting the ideal temperature control mode corresponding to achieving high-quality finished products under different process conditions. In the practical application stage, the control system inputs the current non-temperature process parameters into the model, generates recommended temperature sequences for each hot-pressing position in real time, and dynamically adjusts the output of the heating device to guide the current production process towards historical high-quality levels.

[0115] For example, during the cup wall stretching stage, the model may output a temperature sequence that first rises and then slowly falls to balance material flowability and shaping stability; while during the edge curling stage, it may output a sequence that rapidly heats up to the peak temperature and then holds the temperature to ensure adhesive strength.

[0116] The recommended temperature evolution sequence is used as an upper-level control command and input to the bottom-level PID controller or fuzzy controller of the paper cup hot pressing equipment, which is then converted into a heating power adjustment signal to achieve precise temperature control.

[0117] An adaptive control unit is used to adaptively control the temperature of paper cup hot pressing during real-time production operations based on a temperature distribution prediction model.

[0118] This embodiment constructs a paper cup production record with discrete hot-pressing positions as spatial nodes, realizing the spatiotemporal alignment acquisition of multiple parameters on the mold surface. This allows non-temperature process parameters, temperature control parameters, and finished product quality parameters to form a spatiotemporal graph structure under a unified spatial index and timestamp. This provides cross-parameter correlation based on physical position matching for adaptive control of paper cup hot-pressing temperature, enabling the process behavior and molding quality of each hot-pressing area to correspond point by point.

[0119] Example 2: Referring to Figures 3 to 7, the technical solution of Example 2 differs from that of Example 1 in that it discloses the specific implementation methods of each application unit in Example 1.

[0120] In a preferred embodiment of this invention, the calculation steps of the threshold calculation subunit are as follows:

[0121] B1-1. Based on N paper cup production records within the production cycle, anchor the finished product quality parameters of each paper cup production record;

[0122] B1-2. Based on the finished product quality parameters of each paper cup production record, calculate the quality score of N paper cup production records;

[0123] In this embodiment, the quality score is determined based on the degree to which the finished product quality parameters conform to the preset quality standards. For each paper cup production record, a weighted comprehensive score is calculated based on the compliance of parameters such as thickness, strength, and dimensional deviations. A higher score indicates better molding quality.

[0124] B1-3. Calculate the dynamic quality threshold within the production cycle based on the quality scores of N paper cup production records.

[0125] This embodiment calculates a dynamic quality threshold by independently scoring the quality according to the production cycle, so that the selection of high-quality production records can be adaptively adjusted according to the current process fluctuations. This avoids the sample bias caused by changes in raw material batches or equipment status drift when the fixed threshold is used, and ensures that high-quality production records always represent the relatively optimal process level in the current cycle.

[0126] Furthermore, the calculation step B3 specifically includes:

[0127] B1-3-1. Sort the quality scores of N paper cup production records in ascending order to obtain an ordered quality score sequence.

[0128] B1-3-2. Based on a preset high-quality sample ratio, anchor the target quantile in the ordered quality scoring sequence; wherein, the high-quality sample ratio is 20% to 30%;

[0129] B1-3-3, Extract the quality score corresponding to the target quantile as the dynamic quality threshold within the production cycle.

[0130] For example, if there are N=100 paper cup production records in a certain production cycle, and the preset high-quality sample ratio is 30%, then the target quantile is the 70th percentile (i.e., the top 30% are high quality). After sorting the 100 quality scores, the score value at the 70th percentile (e.g., 87.5 points) is taken as the dynamic quality threshold for that cycle. All records with a score ≥87.5 are judged as high-quality production records.

[0131] This embodiment uses the target quantile of the quality score after sorting as the dynamic quality threshold, so that the screening criterion for each production cycle is directly determined by its internal quality distribution pattern, ensuring that high-quality production records are always at the forefront of the actual output process level in that cycle.

[0132] In a preferred embodiment of this invention, the screening steps of the high-quality screening subunit are as follows:

[0133] B3-1. Based on N paper cup production records for each production cycle, determine M dynamic quality thresholds;

[0134] B3-2. Perform K-value clustering on the M dynamic quality thresholds to obtain K threshold clusters; where each threshold cluster contains several dynamic quality thresholds.

[0135] B3-3. Calculate the threshold variance of each threshold cluster based on the several dynamic quality thresholds contained in each threshold cluster.

[0136] The threshold variance represents the degree of dispersion between the dynamic quality thresholds of each production cycle within the cluster, reflecting the drastic level of quality stability changes in the corresponding process interval during continuous production. The larger the variance, the more significant the fluctuation of the high-quality screening criteria set during the cycle, and the greater the potential disturbance in the corresponding production conditions, such as raw material batch switching, changes in environmental temperature and humidity, or equipment status transitions.

[0137] B3-4. Among the K threshold clusters, select the threshold cluster corresponding to the largest threshold variance and denote it as the target threshold cluster.

[0138] In this embodiment, the cluster corresponding to the maximum threshold variance is selected to focus on production cycles with drastic quality fluctuations. These production cycles often correspond to actual operating conditions with frequent process disturbances. Production records that still exceed the dynamic threshold in these cycles indicate that their forming process has strong anti-interference capabilities. Based on these paper cup production records, the recommended temperature evolution sequence output by the model can be made more environmentally adaptable.

[0139] B3-5. Calculate the average threshold of all dynamic quality thresholds in the target threshold cluster, and denote it as the target threshold.

[0140] The target threshold represents the average quality level corresponding to the process interval with the most significant quality fluctuations across multiple production cycles, reflecting the minimum quality standard required to achieve better molding results within this type of cycle.

[0141] B3-6. From N×M paper cup production records across M production cycles, select P paper cup production records whose quality scores exceed the target threshold and define them as high-quality production records.

[0142] This embodiment determines the cross-cycle target threshold by using the cluster corresponding to the maximum threshold variance, so that the selected high-quality production records are concentrated from production stages with significant process disturbances. This ensures that the temperature control data used for training includes the response behavior to actual disturbance factors such as raw material switching and environmental changes, thereby enhancing the adaptability of the recommended temperature evolution sequence under dynamic production line conditions.

[0143] Furthermore, the screening step B3-2 specifically includes:

[0144] B3-2-1. From M dynamic quality thresholds, randomly select K dynamic quality thresholds as the first round cluster centers;

[0145] B3-2-2. For each dynamic quality threshold, calculate its threshold deviation from the K first-round cluster centers and assign it to the first-round cluster center with the smallest threshold deviation.

[0146] B3-2-3. Traverse all dynamic quality thresholds until each dynamic quality threshold is assigned to the first-round cluster center with the smallest threshold deviation, forming K initial threshold clusters.

[0147] B3-2-4. Calculate the average threshold of each cluster based on all dynamic quality thresholds within each initial threshold cluster, and anchor the average threshold as the cluster center for the next round.

[0148] B3-2-5. Calculate the threshold deviation between the cluster centers of the next round and the cluster centers of the first round;

[0149] B3-2-6. If the threshold deviation is greater than the preset threshold, then all dynamic quality thresholds will be iteratively allocated based on the next round of cluster centers.

[0150] B3-2-7. Repeat the iterative allocation until all threshold deviations are less than the preset threshold, to obtain K threshold clusters;

[0151] B3-2-8. Perform the iterative allocation on each of the M dynamic quality thresholds to obtain K threshold clusters.

[0152] This embodiment updates the cluster centers iteratively and determines the completion of clustering based on convergence conditions, thereby stabilizing the grouping results of the dynamic quality threshold at a local optimum and ensuring that the identified significant quality fluctuation intervals truly reflect the process distribution between production cycles.

[0153] In a preferred embodiment of this invention, the construction steps of the sequence construction subunit are as follows:

[0154] B4-1. Select the target hot-pressing location from among the hot-pressing locations in the high-quality production record;

[0155] B4-2. Obtain the non-temperature process parameters of the target hot-pressing position at T production timestamps; where T production timestamps represent consecutive timestamps on the time axis;

[0156] B4-3. Characterize the non-temperature process parameters into a non-temperature process vector with T production timestamps;

[0157] B4-4. Arrange the T non-temperature process vectors according to the time sequence of the T production timestamps to construct the time series of non-temperature parameters for the target hot pressing position.

[0158] B4-5. Obtain the temperature control parameters of the target hot-pressing location at T production timestamps;

[0159] B4-6. Characterize the temperature control parameters into temperature features of T production timestamps;

[0160] B4-7. Arrange the T temperature features according to the time sequence of the T production timestamps to construct a time series of temperature parameters for the target hot pressing location;

[0161] B4-8. Traverse the high-quality production records and repeatedly execute sequence construction until you obtain the non-temperature parameter time series and temperature parameter time series of H hot pressing positions.

[0162] Both the non-temperature parameter time series and the temperature parameter time series are labeled with the three-dimensional coordinates of their hot-pressing locations.

[0163] This embodiment achieves dual alignment of process parameters in terms of spatial location and time step by independently constructing time series of non-temperature and temperature parameters with three-dimensional coordinate labels for each hot pressing position. This enables the graph structure to establish spatial connection relationships based on three-dimensional coordinates and retain the dynamic characteristics of each position changing over time, supporting consistent mapping between nodes and connecting edges in the spatiotemporal graph structure.

[0164] In a preferred embodiment of this invention, the construction steps of the spatiotemporal graph generation subunit are as follows:

[0165] B5-1. Define the three-dimensional coordinates of the hot-pressing position as nodes, define the non-temperature parameter time series as node features, and define the temperature parameter time series as node labels, thus forming H basic nodes.

[0166] B5-2. Based on the three-dimensional coordinates of the bottom of the paper cup, divide the H basic nodes into cup bottom nodes and cup body nodes.

[0167] B5-3. In the cup bottom nodes, calculate the Euclidean distance between any two cup bottom nodes based on the three-dimensional coordinates of the cup bottom nodes;

[0168] B5-4. If the Euclidean distance is less than the set threshold, a connecting edge is established between the two cup bottom nodes, and the Euclidean distance is inversely proportional to the first heat conduction weight of the connecting edge to obtain the cup bottom diagram structure.

[0169] The formula for calculating the first heat conduction weight is:

[0170] ;

[0171] in, This represents the first heat conduction weight between node i and node j. This represents the Euclidean distance between two nodes. This represents a preset, extremely small positive number, such as 10. -6 This prevents the calculation from being divided by 0.

[0172] B5-5. In the cup body nodes, based on the three-dimensional coordinates of the cup body nodes and the cup body curvature corresponding to the three-dimensional coordinates, calculate the weighted geodesic distance between any two cup body nodes.

[0173] Specifically, the weighted geodetic distance is calculated through the following steps:

[0174] B5-5-1. Based on the 3D CAD model of the mold, calculate the geodesic distance between two nodes using Dijkstra's algorithm or the Fast Marching Method. That is, the shortest path length along the mold surface;

[0175] B5-5-2. Obtain the principal curvature values ​​at nodes i and j.

[0176] B5-5-3. Calculate the mean curvature weighting factor. :

[0177] ;

[0178] Wherein, λ is the curvature influence coefficient (which can be set to 0.1 to 0.5), used to adjust the inhibitory or enhancing effect of curvature on heat conduction. S6-5-4, the weighted geodesic distance is defined as: the average curvature weighting factor. Geodetic distance The product of.

[0179] B5-5-4. If the weighted geodesic distance is less than the set threshold, a directed edge is established between the two cup body nodes, and the geodesic distance is inversely proportional to the second heat conduction weight of the connecting edge to obtain the cup body diagram structure.

[0180] For example, the direction of the directed edge points from the node with smaller curvature to the node with larger curvature, indicating that heat preferentially diffuses to the region with higher curvature.

[0181] The formula for calculating the second heat conduction weight is:

[0182] ;

[0183] in, This represents the second heat conduction weight between node i and node j. This represents the weighted geodesic distance between two nodes.

[0184] B5-5-5. Traverse the H hot-pressing positions in the high-quality production record, merge the cup bottom diagram structure and the cup body diagram structure into a graph structure, and generate a spatiotemporal graph structure of the H hot-pressing positions within the paper cup hot-pressing cycle.

[0185] It should be noted that the aforementioned graph structure fusion represents: based on the three-dimensional coordinates of each hot-pressing position on the mold surface, the nodes in the cup bottom graph structure and the cup body graph structure are uniformly organized to construct a global spatial graph structure covering the cup bottom, cup wall, and rolled edge areas; wherein:

[0186] ① Undirected connecting edges based on Euclidean distance are retained between nodes at the bottom of the cup;

[0187] ② Directed connections based on weighted geodesic distance are maintained between nodes on the cup body;

[0188] ③ In the area where the bottom of the cup meets the body of the cup, if the spatial distance between the nodes of the bottom of the cup and the nodes of the body of the cup is less than the set threshold, a cross-regional connection edge is established, and the connection weight is calculated by weighted average of Euclidean distance and geodesic distance.

[0189] The global spatial diagram structure reflects the spatial adjacency and heat conduction relationship between various hot-pressing positions on the mold surface.

[0190] Furthermore, the spatiotemporal graph structure is expanded along the time dimension based on the global spatial graph: for each time step, the same graph topology is copied, and the node features are the non-temperature parameter time series and temperature parameter time series of each hot pressing position at the production timestamp; the edge structure remains unchanged, thus forming a spatiotemporal graph with a fixed spatial topology and dynamic node features.

[0191] This embodiment constructs differentiated connection relationships between the cup bottom and the cup body based on Euclidean distance and weighted geodesic distance, respectively, and establishes cross-regional connection edges in the boundary area. This allows the global spatial graph structure to accurately reflect the differences in heat conduction paths between the planar and curved areas of the mold surface, while retaining the influence of curvature on the direction of heat diffusion, thus enhancing the consistency between the connection weights between nodes and the heat conduction behavior in the spatiotemporal graph.

[0192] In a preferred embodiment of this invention, the control steps of the adaptive control unit are as follows:

[0193] C1. Real-time acquisition of non-temperature process parameters for each hot-pressing position within a standard time period;

[0194] C2. Construct a real-time spatiotemporal diagram structure from the non-temperature process parameters of each hot pressing position over a standard duration;

[0195] C3. Input the real-time spatiotemporal graph structure into the temperature distribution prediction model and output the recommended temperature evolution sequence of each hot-pressed region within the future time window;

[0196] C4. Based on the recommended temperature evolution sequence of each hot-pressing region within a future time window, adaptive control of the hot-pressing temperature of paper cups is performed.

[0197] The adaptive control refers to the recommended temperature evolution sequence of each hot-pressing region within a future time window based on the output of the temperature distribution prediction model. The heating power of each hot-pressing position is adjusted in real time through PID controllers or fuzzy controllers to make the actual temperature change trajectory as close as possible to the recommended temperature sequence. At the same time, the system continuously monitors the actual temperature data and compares it with the recommended temperature. The model output is corrected by error feedback, thereby forming a closed-loop control loop to ensure precise temperature control during the hot-pressing process of paper cups.

[0198] This embodiment uses the recommended temperature evolution sequence as a dynamic target input to the underlying controller and combines it with error feedback from the actual temperature. This allows the heating power to be adjusted according to the recommended heating path predicted by the model. This path is generated in real time by the temperature distribution prediction model based on the current non-temperature process parameters. It reflects the optimal heating sequence required to achieve high-quality molding at different process stages. It can coordinate the heat energy supply sequence path of each stage of material softening, stretching and molding and cooling, avoid uneven wall thickness and edge cracking caused by temperature lag or overshoot, and ensure the molding integrity of paper cups in complex geometric areas.

[0199] In the embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; multiple units or components may be combined or integrated into another system, or some features may be omitted or not performed. Furthermore, the mutual couplings or direct couplings or communication connections shown or discussed may be indirect couplings or communication connections between apparatuses or units through some interfaces, and may be electrical, mechanical, or other forms.

[0200] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A paper cup hot-pressing temperature adaptive control system based on a deep learning model, characterized in that, include: Record definition unit, used to define paper cup production records; The graph construction unit is used to construct a spatiotemporal graph structure of the paper cup thermoforming process based on the paper cup production records. The graph construction unit includes: a threshold calculation subunit, used to select N paper cup production records within a production cycle in the storage database of paper cup production records, and calculate the dynamic quality threshold within the production cycle; a cycle selection subunit, used to sequentially traverse M production cycles in the storage database based on a preset time window and sliding step size, and select N×M paper cup production records from the M production cycles; a high-quality screening subunit, used to screen P high-quality production records from the N×M paper cup production records in the M production cycles based on the dynamic quality threshold; the screening steps of the high-quality screening subunit include: screening P paper cup production records whose quality scores exceed the target threshold from the N×M paper cup production records in the M production cycles, defining them as high-quality production records; and a sequence construction subunit, used to construct a sequence within each high-quality production record. The system constructs non-temperature parameter time series and temperature parameter time series for H hot-pressing positions; a spatiotemporal graph generation subunit is used to construct a graph structure from the non-temperature parameter time series and temperature parameter time series of the H hot-pressing positions, generating a spatiotemporal graph structure composed of the H hot-pressing positions; a graph training unit is used to traverse P high-quality production records, repeatedly execute the construction of spatiotemporal graph structures until P spatiotemporal graph structures are obtained, and use the P spatiotemporal graph structures as training samples for the time-series graph neural model to iteratively train the temperature distribution prediction model; the training objective of the temperature distribution prediction model is defined as: under the given non-temperature parameter time series of H hot-pressing positions, predict the recommended temperature evolution sequence of each hot-pressing position within a future time window; an adaptive control unit is used to adaptively control the paper cup hot-pressing temperature in real-time production operations based on the temperature distribution prediction model.

2. The paper cup hot-pressing temperature adaptive control system based on a deep learning model according to claim 1, characterized in that, The calculation steps of the threshold calculation subunit are as follows: B1-1, anchor the finished product quality parameters of each paper cup production record based on the N paper cup production records in the production cycle; B1-2, calculate the quality score of the N paper cup production records based on the finished product quality parameters of each paper cup production record. B1-3. Calculate the dynamic quality threshold within the production cycle based on the quality scores of N paper cup production records.

3. The paper cup hot-pressing temperature adaptive control system based on a deep learning model according to claim 2, characterized in that, Calculating the dynamic quality threshold within the production cycle includes: B1-3-1, sorting the quality scores of N paper cup production records in ascending order to obtain an ordered quality score sequence; B1-3-2, anchoring the target quantile in the ordered quality score sequence according to a preset high-quality sample ratio; B1-3-3, extracting the quality score corresponding to the target quantile as the dynamic quality threshold within the production cycle.

4. The paper cup hot-pressing temperature adaptive control system based on a deep learning model according to claim 3, characterized in that, The screening steps of the high-quality screening subunit further include: B3-1, determining M dynamic quality thresholds based on N paper cup production records for each production cycle; B3-2, performing K-value clustering on the M dynamic quality thresholds to obtain K threshold clusters; wherein each threshold cluster contains several dynamic quality thresholds; B3-3, calculating the threshold variance of each threshold cluster based on the several dynamic quality thresholds contained in each threshold cluster; B3-4, selecting the threshold cluster corresponding to the largest threshold variance among the K threshold clusters, and denoting it as the target threshold cluster; B3-5, calculating the average threshold of all dynamic quality thresholds in the target threshold cluster, and denoting it as the target threshold.

5. The paper cup hot-pressing temperature adaptive control system based on a deep learning model according to claim 4, characterized in that, K-value clustering is performed on the M dynamic quality thresholds to obtain K threshold clusters, including: B3-2-1, randomly selecting K dynamic quality thresholds from the M dynamic quality thresholds as the first-round cluster centers; B3-2-2, for each dynamic quality threshold, calculating its threshold deviation from the K first-round cluster centers, and assigning it to the first-round cluster center with the smallest threshold deviation; B3-2-3, traversing all dynamic quality thresholds until each dynamic quality threshold is assigned to the first-round cluster center with the smallest threshold deviation, forming K initial threshold clusters; B3-2-4, based on each B3-2-5. Calculate the average threshold of all dynamic quality thresholds within the initial threshold cluster and anchor the average threshold as the cluster center for the next round; B3-2-6. Calculate the threshold deviation between the cluster center for the next round and the cluster center for the first round; B3-2-7. If the threshold deviation is greater than a preset threshold, iteratively allocate all dynamic quality thresholds based on the cluster center for the next round; B3-2-8. Repeat the iterative allocation until all threshold deviations are less than the preset threshold, resulting in K threshold clusters; B3-2-9. Perform the iterative allocation on M dynamic quality thresholds respectively to obtain K threshold clusters.

6. The paper cup hot-pressing temperature adaptive control system based on a deep learning model according to claim 5, characterized in that, The construction steps of the sequence construction subunit are as follows: B4-1, Select the target hot-pressing position from among the hot-pressing positions in the high-quality production record; B4-2, Obtain the non-temperature process parameters of the target hot-pressing position at T production timestamps; where T production timestamps represent consecutive timestamps on the time axis; B4-3, Characterize the non-temperature process parameters into non-temperature process vectors at T production timestamps; B4-4, Arrange the T non-temperature process vectors according to the time order of the T production timestamps to construct the non-temperature parameter time series of the target hot-pressing position; B4-5. Obtain the temperature control parameters of the target hot-pressing location at T production timestamps; B4-6, characterize the temperature control parameters into temperature features at T production timestamps; B4-7, arrange the T temperature features in chronological order according to the T production timestamps to construct the temperature parameter time series of the target hot-pressing location; B4-8, traverse high-quality production records and repeat the sequence construction until H non-temperature parameter time series and temperature parameter time series of hot-pressing locations are obtained; where both the non-temperature parameter time series and the temperature parameter time series are labeled with the three-dimensional coordinates of their hot-pressing locations.

7. The paper cup hot-pressing temperature adaptive control system based on a deep learning model according to claim 6, characterized in that, The construction steps of the spatiotemporal graph generation subunit are as follows: B5-1, define the three-dimensional coordinates of the hot-pressing position as nodes, define the non-temperature parameter time series as node features, and define the temperature parameter time series as node labels, forming H basic nodes; B5-2, among the H basic nodes, based on the three-dimensional coordinates of the bottom of the paper cup, divide the H basic nodes into cup bottom nodes and cup body nodes; B5-3, in the cup bottom nodes, calculate the Euclidean distance between any two cup bottom nodes based on the three-dimensional coordinates of the cup bottom nodes; B5-4, if the Euclidean distance is less than a set threshold, establish a connecting edge between the two cup bottom nodes, and inversely proportional the Euclidean distance to the threshold value. The first heat conduction weight of the connecting edge is used to obtain the cup bottom diagram structure; B5-5, in the cup body nodes, based on the three-dimensional coordinates of the cup body nodes and the corresponding cup body curvature, the weighted geodesic distance between any two cup body nodes is calculated; B5-6, if the weighted geodesic distance is less than a set threshold, a directed edge is established between the two cup body nodes, and the weighted geodesic distance is inversely proportional to the second heat conduction weight of the connecting edge to obtain the cup body diagram structure; B5-7, traverse the H hot pressing positions in the high-quality production record, merge the cup bottom diagram structure and the cup body diagram structure into a graph structure to generate a spatiotemporal graph structure of the H hot pressing positions within the paper cup hot pressing forming cycle.

8. The paper cup hot-pressing temperature adaptive control system based on a deep learning model according to claim 7, characterized in that, The control steps of the adaptive control unit are as follows: C1, real-time acquisition of non-temperature process parameters of each hot pressing position within a standard time period; C2, constructing a real-time spatiotemporal diagram structure from the non-temperature process parameters of each hot pressing position within the standard time period. C3. Input the real-time spatiotemporal graph structure into the temperature distribution prediction model and output the recommended temperature evolution sequence of each hot-pressing region within the future time window; C4. Based on the recommended temperature evolution sequence of each hot-pressing region within the future time window, perform adaptive control of the paper cup hot-pressing forming temperature.

Images