Process parameter adaptive adjustment system for ampoule bottling process

Abstract

The application discloses a process parameter self-adaptive adjusting system for an ampoule manufacturing process, and relates to the field of automatic control, which comprises a cutting parameter determining module, a cross-section quality multi-parameter online detecting module, a round mouth process parameter determining module, a round mouth detecting module and a parameter closed loop adjusting module connected in sequence. The round mouth process parameter is dynamically determined by using a multi-branch physical information neural network, the network is optimized by cooperation of thermodynamics, fluid mechanics and quality prediction branches, and a physical checking mechanism is embedded to ensure that the parameter is reasonable. The system guides accurate parameter optimization through a quality potential gradient, and realizes multi-objective balance such as breaking force and smoothness. The application can dynamically compensate working condition fluctuation according to the cross-section parameter, improve the quality consistency of the bottle mouth, reduce glass debris, and guarantee the safety of medicine packaging.

Description

Technical Field

[0001] This invention relates to the field of automation control, and in particular to an adaptive adjustment system for process parameters in the ampoule bottle making process. Background Technology

[0002] In ampoule production, cutting and mouth rounding are critical processes that determine the quality of the bottle mouth and the safety of drug filling. Current processes suffer from two major drawbacks: 1) Cutting and mouth rounding are controlled independently, with parameter settings relying on experience and unable to adaptively adjust to the microscopic differences in the incoming glass tubes; 2) Quality inspection is lagging, with only random checks possible after mouth rounding, failing to achieve real-time closed-loop process control, leading to batch defects such as debris, microcracks, and uneven mouth rounding.

[0003] Existing technologies provide general methods for bottle manufacturing, but they do not solve the problem of dynamic parameter coupling optimization based on real-time cross-sectional quality, nor the closed-loop adaptive control architecture. Therefore, there is an urgent need for an adaptive adjustment method that can perceive in real time, make intelligent decisions, and precisely coordinate. Summary of the Invention

[0004] To address the technical problems of automation control in the prior art, this invention provides an adaptive adjustment system for process parameters in the ampoule bottle making process.

[0005] This invention is achieved through the following technical solution:

[0006] The present invention also provides an adaptive adjustment system for process parameters in the ampoule manufacturing process, comprising:

[0007] The module includes: cutting parameter determination module, cross-section quality multi-parameter online detection module, round mouth process parameter determination module, round mouth detection module, and parameter closed-loop adjustment module.

[0008] The cutting parameter determination module assigns appropriate cutting parameters to the glass tube to be cut.

[0009] The online multi-parameter detection module for cross-section quality measures the cut cross-section to obtain the cross-section roughness, perpendicularity of the cross-section to the glass tube axis, maximum crack depth, and width of the heat-affected zone.

[0010] The circular inlet process parameter determination module dynamically determines the circular inlet process parameters based on the real-time detected cross-sectional parameters and a hierarchical multi-branch physical information neural network.

[0011] The round-mouth inspection module performs online inspection of the finished product after rounding to form a quality evaluation.

[0012] The parameter closed-loop adjustment module uses the data detected by the circular port detection module to continuously adjust the model parameters in the system.

[0013] Furthermore, the cross-sectional roughness is obtained based on the acquisition of the cross-sectional profile by a white light confocal displacement sensor, which obtains multiple contour lines of the entire cross-section, forming a cross-sectional point cloud. The arithmetic mean deviation of all point clouds to the fitted plane is calculated to obtain the cross-sectional roughness.

[0014] Furthermore, the hierarchical multi-branch physical information neural network includes a shared backbone network, a first branch network, a second branch network, a third branch network, a physical information exchange and collaboration network, and a physical constraint output layer.

[0015] Furthermore, the first branch network is a thermodynamic branch, and the input is... ;in, For cross-sectional feature vectors, For the context from the fluid branch, The output is a temperature curve, representing the mass potential gradient from the mass branch. and total heating time The second branch network is a fluid branch, with the input being... ;in, In the context of the thermodynamics branch, the output is the fuel gas ratio. and rotational speed The third branch network is the quality prediction branch, and its input is... The output includes a predicted quality vector. and quality potential gradient .

[0016] Furthermore, the first branch network also includes an attention-enhanced LSTM unit. In the cell state update of the LSTM, a cross-branch attention gate is added. This gate calculates the correlation between the current LSTM hidden state and the fluid branch context, and determines how much information to absorb from the fluid branch context to correct the current thermal state prediction.

[0017] Furthermore, the first branch network also includes a micro-physics solver, which is used to calculate the theoretical thermal stress distribution under the temperature curve. If it exceeds the safety threshold, a gradient signal is generated to correct the temperature curve in reverse.

[0018] Furthermore, the second branch network is based on a residual decision network, including a basic decision path and a correction path. The basic decision path directly generates the initial action from the cross-sectional feature vector, while the correction path integrates the context of the thermodynamic branch and the mass potential gradient to generate a correction quantity. The output is the sum of the initial action and the correction quantity.

[0019] Furthermore, the quality potential gradient in the third branch network is calculated as follows:

[0020] ;

[0021] Where ΔQ is the target mass With prediction quality The difference, where J is the Jacobian matrix.

[0022] Furthermore, the physical information exchange and collaboration network includes generating a consistency signal and calculating the degree of matching between the theoretically required heat flux intensity value and the theoretical flame temperature value, i.e., the thermodynamic matching degree.

[0023] Furthermore, the expression for the loss function of the hierarchical multi-branch physical information neural network is as follows:

[0024] ;

[0025] in, The main prediction loss is the mean square error between the model's final output and the actual optimal process parameters. Loss due to physical laws; This results in a loss of coordination and consistency. To aid in loss, we characterize the difference between the output of the quality prediction branch and the actual quality.

[0026] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0027] By integrating white light confocal microscopy, high-frequency ultrasound, and short-wave infrared thermal imagers, a multi-sensor fusion detection platform was constructed, enabling the first-ever simultaneous acquisition of complete digital images of cross-sectional roughness, perpendicularity, microcrack depth, and heat-affected zone width within milliseconds. This overcomes the limitations and delays of traditional manual sampling or single-sensor detection.

[0028] The innovative Multi-Branch Physical Information Neural Network (MBPINN) architecture, through the specialized division of labor and real-time information exchange among thermodynamics, fluid mechanics, and mass prediction branches, can dynamically output a self-consistent and collaborative optimal set of circular parameters, achieving global optimization. Furthermore, a differentiable micro-physical solver is embedded in the neural network to perform online physical verification of the output temperature curves, etc. This fundamentally eliminates parameter outputs that violate thermodynamic laws and could lead to equipment damage or safety accidents, greatly improving the system's reliability and robustness.

[0029] This invention solves the optimization challenge of balancing multiple quality objectives: the quality prediction branch not only predicts quality but also innovatively outputs a quality potential gradient. This gradient serves as a precise navigation signal, clearly instructing the thermodynamics and fluid mechanics branches on how to adjust parameters to comprehensively improve indicators such as breaking force and smoothness, thus realizing a shift from blind trial and error to goal-oriented precise optimization. Attached Figure Description

[0030] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0031] Figure 1 This is a schematic diagram of the control method used in an adaptive adjustment system for process parameters in an ampoule bottle making process, according to an embodiment of this application. Detailed Implementation

[0032] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0033] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0034] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. The illustrations only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the shape, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0035] See Figure 1 An adaptive adjustment system for process parameters in ampoule bottle making includes:

[0036] The module includes a cutting parameter determination module, a multi-parameter online detection module for cross-section quality, a round mouth process parameter determination module, a round mouth detection module, and a parameter closed-loop adjustment module.

[0037] S1. The cutting parameter determination module assigns appropriate cutting parameters to the glass tube to be cut, serving as the starting point for process execution.

[0038] Specifically, this invention employs a cutting parameter determination model built based on a Random Forest regression algorithm to determine the cutting process parameters. The cutting process parameters, i.e., the model output, include the dynamic pressure of the cutting wheel, the linear velocity of the glass tube rotation, and the cooling intensity of the cold air after the cutting point. The model input parameters are the feature vectors of the glass tube being cut, including tube diameter, wall thickness, relative hardness scalar, and glass tube temperature.

[0039] Specifically, the input parameters are: .in, Pipe diameter (mm). The wall thickness is in mm. This is a scalar measure of relative hardness. The temperature of the glass tube before it enters the cutting station (°C).

[0040] Output parameters: .in, The dynamic pressure of the cutter wheel (N). The linear velocity of the glass tube rotation (mm / s) is given. The cooling intensity of the cold air after the cutting point (L / min).

[0041] S2, the online multi-parameter cross-section quality detection module, is a key sensing component connecting physical cutting and intelligent decision-making. Its goal is to complete multi-dimensional, non-destructive, and high-precision measurements of the freshly cut cross-section within milliseconds, and to transform the physical features into a precise digital feature vector S that can be processed by artificial intelligence models.

[0042] To achieve simultaneous capture of four core parameters (section roughness, perpendicularity of the section to the glass tube axis, maximum crack depth, and width of the heat-affected zone), this invention constructs an online detection platform with multiple sensors working in collaboration, which follows the cutting station.

[0043] The platform consists of a high-precision rotary clamping unit, a multi-sensor array, and a data acquisition card.

[0044] The high-precision rotary clamping unit clamps the ampoule at a constant speed and rotates it around its axis, ensuring that the sensor can scan the entire cross-sectional circumference.

[0045] The multi-sensor array is arranged sequentially along the axial direction of the bottle, completing all measurements within one rotation of the bottle. Specifically, the multi-sensor array includes a white light confocal displacement sensor for detecting cross-sectional roughness and the perpendicularity of the cross-section to the glass tube axis, a high-frequency ultrasonic probe for detecting crack depth, and a high-speed short-wave infrared thermal imager for detecting the width of the heat-affected zone.

[0046] The data acquisition card ensures that the acquisition actions of all sensors are strictly synchronized with the rotation angle of the bottle, guaranteeing the spatial alignment of the data.

[0047] Specifically, the measurement method is as follows:

[0048] For measuring cross-sectional roughness and perpendicularity of the cross-section to the glass tube axis, the white light confocal displacement sensor is used because it has nanometer-level longitudinal resolution and strong adaptability to highly reflective surfaces such as glass.

[0049] Specifically, the sensor performs a radial scan along the cross-sectional contour to obtain a height profile curve z(x). The bottle rotates one full turn to obtain N profile lines of the entire cross-section, forming a cross-sectional point cloud.

[0050] Further feature extraction is performed:

[0051] First, a baseline plane is fitted. This involves using a random sampling consensus algorithm to filter out outliers (such as chipped edges and debris) from the cross-sectional point cloud data, and then using the least squares method to fit an ideal plane.

[0052] Surface roughness The calculation involves calculating the arithmetic mean deviation of all point clouds from the fitted plane:

[0053] ;

[0054] Where (A, B, C) are the parameters of the fitted plane, and N is the number of point clouds. Let be the coordinates of the i-th point cloud.

[0055] Perpendicularity of the cross-section to the axis of the glass tube Calculate the angle between the normal vector n = (A, B, -1) of the fitted plane and the direction vector a = (0, 0, 1) of the glass tube axis.

[0056] ;

[0057] For the maximum crack depth A high-frequency ultrasonic probe is used to excite surface waves to propagate along the cross-section. When the wave encounters a crack, some of the sound waves are reflected. By measuring the time difference Δt between the emitted and reflected waves, and combining this with the sound velocity v of the surface wave in the glass, the crack depth can be calculated. This method is particularly sensitive to closed cracks. Simultaneously, orthophotos of the cross-section are acquired using a telecentric microscope. For the microscopic images, a pre-trained U-Net deep learning model is used for pixel-level segmentation to identify all visible crack regions and estimate their opening width and length.

[0058] Spatial matching and correlation are performed between multiple depth peaks detected by ultrasound and the crack locations segmented from the image.

[0059] The maximum crack depth value is output, taking the maximum depth value measured ultrasonically from all successfully matched cracks. This method balances the objectivity of the measurement (ultrasonic physical measurement) and its interpretability (image visual association).

[0060] For the width of the heat-affected zone, a high-speed short-wave infrared thermal imager with a frame rate exceeding 1000 fps is used to capture the transient temperature field at the moment of cutting. The thermal imager continuously records the temperature field video T(x,y,t) of the area near the cut before and after the cutting action is triggered. The temperature-time curve is extracted to find the point of highest temperature. and their corresponding time ,exist In the frame, the temperature distribution curve is plotted radially along the glass tube, centered on the point of highest temperature. Width of the heat-affected zone. Defined as temperature from Decrease to glass transition temperature The radial distance. This characterizes the size of the region where heating causes irreversible changes in the microstructure.

[0061] The measurement data for the four parameters are processed by their respective algorithms and then converged into the data fusion unit. The fusion unit ultimately packages these data into a standardized feature vector:

[0062] ;

[0063] The vector includes a timestamp, preform ID, and an index of the original data snapshot for traceability.

[0064] S3, the round-mouth process parameter determination module is the core of this invention. Based on real-time detected cross-sectional parameters, it dynamically determines the round-mouth process parameters. The core significance of dynamically determining the round-mouth parameters lies in achieving personalized and precise production. The performance parameters of each batch of glass bottles are unique, exhibiting slight differences. Traditional fixed-parameter methods cannot differentiate between batches, leading to uneven quality and the risk of glass breakage. By analyzing cross-sectional characteristics in real time, an optimal set of round-mouth temperature, flame, and motion parameters is dynamically matched to each bottle. This fundamentally eliminates batch defects caused by fixed processes, maximizing the consistency of bottle mouth quality and virtually eliminating glass breakage, ensuring absolute safety in pharmaceutical filling from the source.

[0065] This invention employs a hierarchical, multi-branch Physics-Informed Neural Network (MBPINN). Its core idea is to process subsets of parameters with different physical properties and influencing mechanisms through specialized subnetworks (branches), and then, through a carefully designed physical information exchange module, allow the branches to dynamically collaborate during the decision-making process, ultimately outputting a self-consistent, feasible, and optimal parameter set.

[0066] The network's input parameters are the cross-sectional feature vector S, and its output parameters are the set of process parameters for the circular inlet. .in, This represents the flow ratio of oxygen to fuel gas. The temperature curve is shown. The rotational speed of the ampoule in the flame. This represents the total heating time. The temperature curve includes three key points: preheating temperature, peak melting temperature, and slow cooling temperature.

[0067] The hierarchical multi-branch physical information neural network includes a shared backbone network, a first branch network, a second branch network, a third branch network, a physical information exchange and collaboration network, and a physical constraint output layer.

[0068] S31, Shared backbone network;

[0069] The cross-sectional feature vector is input into a shared backbone network, which consists of a multilayer perceptron. Its function is to learn to map the original cross-sectional features S to a high-dimensional, abstract shared feature space. .

[0070] ;

[0071] MLP stands for Multilayer Perceptron. The set of trainable weight parameters for the shared feature extraction backbone network. It is a feature vector that contains all the abstract information of the cross-sectional state, providing a general and high-level understanding of the working conditions for subsequent branches.

[0072] S32, First Branch Network;

[0073] The first branch is thermodynamics, used to predict parameters directly related to heat input, distribution, and dissipation, i.e., temperature curves. and total heating time .

[0074] The input for the first branch is ;in, For the context from the fluid branch, This represents the mass potential gradient from the mass branch, for example, "To increase breaking force, the effective heat input needs to be increased." The output is a temperature curve. and total heating time .

[0075] Furthermore, fluid branching context It is obtained by connecting a specific hidden layer inside the fluid branch network to a lightweight context-coded subnetwork.

[0076] The network structure of the first branch includes Long Short-Term Memory (LSTM) network units to explicitly model the sequence dependencies of temperature changes over time.

[0077] The network's first layer is not a fully connected layer, but a physical projection layer. It maps the input features to a low-dimensional space composed of physical dimensions, forcing the network to think about problems from a physical perspective.

[0078] Preferably, an attention-enhanced LSTM unit is used, in which a cross-branch attention gate is added to the LSTM cell state update. This gate calculates the current LSTM hidden state and the fluid branch context. The correlation, and based on this, decide from How much information is absorbed to correct the current thermal state prediction? Specifically:

[0079] Step A: Calculate the attention score:

[0080] First, calculate the hidden state of the LSTM at time t-1. With fluid context Correlation of each element in:

[0081] ;

[0082] in, , , and It consists of trainable weight matrices and offsets. It is a scalar fraction, representing The importance of the current time step.

[0083] Step B: Generate attention weights:

[0084] The attention score is converted into a gating value between 0 and 1 using the Sigmoid function. .

[0085] ;

[0086] That is, to decide from The proportion of information absorbed. If → 1 indicates that the current thermal state prediction is highly dependent on flame information; if → 0 indicates to ignore.

[0087] Step C: Generate attention-modulated context vectors:

[0088] ;

[0089] Here Through a linear transformation Projected into the same space as the LSTM hidden state, and weighted... Scaling. It is the fully connected weight matrix obtained through training.

[0090] Step D: Integrating LSTM core computing:

[0091] modulated context vector It is added as additional input to the LSTM cell state update calculation. A common approach is to combine it with the regular input. splicing:

[0092] Input enhancement: Then, Feed the data into a standard LSTM-gated computation to update the cell state. and hidden state .

[0093] This mechanism allows the thermodynamic LSTM to dynamically and selectively reference the current flame state when predicting the temperature at each moment. For example, when predicting the preheating section temperature, The temperature may be high, and the system will primarily refer to the flame's convective heat transfer coefficient to accurately predict the heating rate. When predicting the peak holding temperature, The focus might then shift to the temperature of the adiabatic flame to determine if the heat source is sufficient. This dynamic reference enables coupled modeling of the "heat source-workpiece" system.

[0094] The model also includes a micro-physics solver. After the network outputs the original temperature-time series, it is not directly passed; instead, it is input into a differentiable simplified heat transfer model for verification. This solver calculates the theoretical thermal stress distribution under this temperature curve. If it exceeds a safety threshold, it generates a gradient signal to correct the temperature curve in reverse. Specifically:

[0095] Input the initial temperature curve predicted by the neural network and the known material parameters.

[0096] Output: Physically reliable temperature profiles and physical violation losses. .

[0097] Step A: Establish a simplified model:

[0098] Using a one-dimensional unsteady-state heat conduction equation as the core, it is assumed that heat is mainly conducted radially along the bottle wall:

[0099] ;

[0100] in, Let r be the temperature at radial coordinate r and time t. For thermal diffusivity, For net heat source term, related to temperature and The estimated surface heat flow is related.

[0101] Step B: Discretization and Solution:

[0102] Discretize the time and space, and solve the above equations using differentiable numerical methods such as the finite difference method. The key to this step is that all operations are implemented using differentiable tensor operations.

[0103] Step C: Calculate physical constraint penalties:

[0104] The calculation can obtain the inner and outer wall temperatures of the bottle opening and the radial temperature gradient. and the resulting thermal stress .

[0105] Calculate physical loss terms, such as penalizing excessive thermal stress or too rapid heating rates:

[0106] ;

[0107] ;

[0108] in, , These represent the maximum thermal stress and the maximum radial temperature gradient, respectively.

[0109] ;

[0110] Step D: Forward propagation and gradient backpropagation:

[0111] Forward: The solver receives the temperature curve, solves for the internal temperature field, and calculates... Based on the physically reliable temperature curves of the key points at the bottle opening after the solution, it is also possible to... As a feedback signal.

[0112] Inverse: Total loss during training .in, The loss function for the task is used to ensure that the model is effective. Ensure the model conforms to physical rules.

[0113] The resulting gradient is backpropagated through this differentiable solver to the original input temperature curve, thus guiding the neural network to adjust its output in the direction of reducing physical violations during the next prediction.

[0114] This miniature solver acts as an embedded, differentiable "physics checker." It forces the neural network's output to obey fundamental laws of thermodynamics in addition to being data-driven. This significantly improves the physical reliability and generalization ability of the network's output, especially in extreme conditions not covered by the training data, preventing physically absurd predictions.

[0115] S33, Second Branch Network;

[0116] The second branch is the fluid branch, used to predict parameters affecting flame morphology and material movement, namely the fuel gas ratio. and rotational speed .

[0117] enter: ;in, For the context from the branch of thermodynamics, The quality potential gradient for the quality branch, for example, "To improve smoothness, the flow field uniformity of the glass melt needs to be optimized."

[0118] The network structure is based on residual decision networks, specifically including:

[0119] Conditional normalization layer: A conditional batch normalization layer is added before the first layer, and its scaling and translation parameters are determined by... Dynamic generation. This allows the network to quickly switch to different response modes in response to different thermal schemes.

[0120] Residual Decision Network: Employs a residual block structure. The basic decision path is directly derived from... Generate initial actions Another corrective path combines... and Generate a correction value The output is This enables rapid and precise fine-tuning of fundamental decisions.

[0121] Feasible region activation function: The last layer uses a custom sigmoid function, whose upper and lower bounds are not fixed, but rather depend on... A dynamically calculated safety operation window. For example, it automatically narrows the fuel-gas ratio adjustment range when there is high heat input to prevent overburning.

[0122] The final output is a fuel ratio that adapts to environmental conditions. and rotational speed .

[0123] S34, the third branch network;

[0124] The third branch is the quality prediction branch, and the network input is... The output consists of two parallel output heads. Output head A is the state prediction, outputting a standard prediction quality vector. The output header B is for gradient prediction, which is the core improvement. This header outputs an approximation of a Jacobian matrix J, where J[i, j] characterizes the marginal effect of a small change in the j-th process parameter (such as the fuel gas ratio) on the i-th quality index (such as breaking force). This is achieved by designing dual forward propagation in the network and combining it with automatic differentiation techniques.

[0125] Generate a quality potential gradient vector: The system will target the quality gradient vector. With prediction quality The difference ΔQ, combined with the Jacobian matrix J, is used to calculate the optimal adjustment direction:

[0126] ;

[0127] this It is a vector that explicitly tells the thermodynamic and fluid mechanics branches which parameters should change in order to approach the target.

[0128] Expected output quality quality potential gradient vector .

[0129] The three branches of this invention are not simply parallel structures, but rather constitute an intelligent system that co-evolves through "perception-decision-evaluation". To address the core issues of strong parameter coupling and dynamic changes in quality targets in practical processes, the three branches are now systematically restructured:

[0130] By introducing a gated attention mechanism and a micro-prediction loop, thermodynamics and fluid mechanics branches can exchange information in multiple rounds to achieve optimal coordination between parameters.

[0131] Strengthen the role of the quality prediction branch so that it can not only predict the final quality, but also output a quality potential gradient vector ∇Q, which directly and quantitatively guides the other two branches to adjust their output in which direction.

[0132] By embedding lightweight physics equation solvers within branches, the neural network is ensured to explore within a reasonable physical space, accelerating training and improving the physical reliability of the output parameters.

[0133] S35, Physical Information Exchange and Collaboration; this is the key to the model's intelligent collaboration. It receives intermediate layer features or preliminary outputs from the three branches, performs calculations, and then feeds the results back to the control parameter branch.

[0134] First, a heat-fluid coupling calculation is performed. Using a built-in simplified, differentiable physical calculation unit, a theoretically required heat flux value is calculated based on the temperature curve of the thermodynamic branch. Meanwhile, based on the gas ratio from the hydrodynamic branch, a theoretical flame temperature value, Flametemp, is calculated.

[0135] Further, a consistency signal is generated, and the theoretically required heat flux intensity value is calculated. The degree of matching between the theoretical flame temperature value (Flametemp) and the thermodynamic matching degree. This signal reflects whether the preliminary schemes of the two branches are physically consistent.

[0136] Calculate the target gap signal, and then calculate the output of the quality prediction branch. With the highest quality target quality difference between get.

[0137] Optionally, thermodynamic matching degree and quality difference As additional contextual features, these are fed back to the intermediate layers of the thermodynamics and fluid dynamics branches, respectively. This allows these two branches to instantly perceive the other's intentions and the quality gap of their common goals during subsequent forward propagation, thereby dynamically adjusting their outputs.

[0138] S36, Physical Constraint Output Layer

[0139] The original outputs of each branch are iteratively adjusted by the exchange module before being sent to this layer for final physical constraint encapsulation.

[0140] Input adjusted The processing includes numerical range constraints and temporal relationship constraints.

[0141] The numerical range constraint: The activation function is used to strictly limit the output within the process safety window.

[0142] The temporal relationship constraint ensures the physical rationality of the temperature curve through a constructive function.

[0143] Output the final feasible set of circular nozzle process parameters.

[0144] Next, the loss function is determined and the model is trained. The model adopts a phased training strategy:

[0145] Pre-training: The shared backbone network, the quality prediction branch, and the thermodynamics and fluid dynamics branches are pre-trained separately using historical data. During this stage, the exchange module is not in operation.

[0146] Joint fine-tuning: Enable the full hierarchical multi-branch physical information neural network MBPINN, including the switching module, for end-to-end joint training.

[0147] The total loss function consists of four parts:

[0148] ;

[0149] in, The main prediction loss is the mean square error between the model's final output and the actual optimal process parameters. The violation of rigid physical rules is punished as a result of the loss of physical laws. To mitigate the loss of consensus, the consensus signal calculated by the switching module is encouraged to approach 0, which means that the outputs of the two parameter branches are forced to be physically self-consistent. To assist in the loss, the loss between the output of the quality prediction branch and the actual quality is characterized, which is used to stabilize the learning of shared features and guide branch collaboration.

[0150] The optimal set of process parameters for the circular inlet corresponding to a specific cross-section is finally obtained. This output directly drives the circular inlet equipment to execute.

[0151] S4, the round mouth detection module, is used to perform rapid online detection on the finished product after round mouth, forming a final quality evaluation as the basis for closed-loop feedback.

[0152] The quality of the round mouth is obtained from simulated breaking force, the smoothness of the molten surface of the bottle mouth, and a vision-based defect score. Specifically, the round mouth quality vector... ,in, To simulate breaking force, For the smoothness of the molten surface at the bottle neck, Visual-based defect scoring.

[0153] Optional, The force was measured by a micro-force sensor during a simulated breaking action by a robotic arm. Images of the bottle opening are captured using a high-speed industrial camera, and then analyzed and scored using a convolutional neural network (CNN) image regression and classification model.

[0154] S5, parameter closed-loop adjustment module;

[0155] The data detected by the circular port detection module is used to continuously optimize the model in the system, thereby achieving self-evolution.

[0156] Specifically, reinforcement learning-based model optimization engines include:

[0157] Data storage: Complete production data for each ampoule is stored in the process knowledge base.

[0158] Reward Calculation: Calculate a comprehensive reward value for each data point. Where f, g, and h are standardization functions, w is the weight, and a higher reward value R indicates better overall quality.

[0159] Model Update: New data from the knowledge base is periodically used to fine-tune the model network parameters using a policy gradient method, aiming to maximize the average reward R. Simultaneously, the updated model is used to back-evaluate the feature vectors and cutting parameters of different glass tubes historically. The ultimate quality potential that can be generated by combination, thereby generating new optimized samples. The cutting parameters are used to determine the model for retraining step S1.

[0160] The updated cutting process parameters and round-mouth process parameters are output to determine the model. The updated model will be used in subsequent production, forming a continuous improvement closed loop of production-learning-optimization-reproduction.

[0161] In this implementation, the following is achieved:

[0162] By integrating white light confocal microscopy, high-frequency ultrasound, and short-wave infrared thermal imagers, a multi-sensor fusion detection platform was constructed, enabling the first-ever simultaneous acquisition of complete digital images of cross-sectional roughness, perpendicularity, microcrack depth, and heat-affected zone width within milliseconds. This overcomes the limitations and delays of traditional manual sampling or single-sensor detection.

[0163] The innovative Multi-Branch Physical Information Neural Network (MBPINN) architecture, through the specialized division of labor and real-time information exchange among thermodynamics, fluid mechanics, and mass prediction branches, can dynamically output a self-consistent and collaborative optimal set of circular parameters, achieving global optimization. Furthermore, a differentiable micro-physical solver is embedded in the neural network to perform online physical verification of the output temperature curves, etc. This fundamentally eliminates parameter outputs that violate thermodynamic laws and could lead to equipment damage or safety accidents, greatly improving the system's reliability and robustness.

[0164] This invention solves the optimization challenge of balancing multiple quality objectives: the quality prediction branch not only predicts quality but also innovatively outputs a quality potential gradient. This gradient serves as a precise navigation signal, clearly instructing the thermodynamics and fluid mechanics branches on how to adjust parameters to comprehensively improve indicators such as breaking force and smoothness, thus realizing a shift from blind trial and error to goal-oriented precise optimization.

[0165] By dynamically adjusting the cross-sectional parameters, the system can proactively compensate for the impact of fluctuations in glass raw materials and changes in equipment status. This is expected to improve batch consistency of key quality indicators at the bottle opening by more than 50% and almost eliminate glass fragments caused by process mismatch, thus ensuring the safety of pharmaceutical packaging from the source.

[0166] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A process parameter adaptive adjustment system for ampoule bottle making, characterized in that, include: The module includes: cutting parameter determination module, cross-section quality multi-parameter online detection module, round mouth process parameter determination module, round mouth detection module, and parameter closed-loop adjustment module. The cutting parameter determination module assigns cutting parameters to the glass tube to be cut. The online multi-parameter detection module for cross-section quality measures the cut cross-section to obtain the cross-section roughness, perpendicularity of the cross-section to the glass tube axis, maximum crack depth, and width of the heat-affected zone. The circular inlet process parameter determination module dynamically determines the circular inlet process parameters based on the real-time detected cross-sectional parameters and a hierarchical multi-branch physical information neural network. The round-mouth inspection module performs online inspection of the finished product after rounding to form a quality evaluation. The parameter closed-loop adjustment module uses the data detected by the circular port detection module to continuously adjust the model parameters in the system; The hierarchical multi-branch physical information neural network includes a shared backbone network, a first branch network, a second branch network, a third branch network, a physical information exchange and collaboration network, and a physical constraint output layer; The first branch is the thermodynamic branch, used to predict parameters directly related to heat input, distribution, and dissipation, i.e., temperature curves. and total heating time The input to the first branch is ;in, The cross-sectional feature vector; For the context from the fluid branch, The quality potential gradient comes from the quality branch; the network structure of the first branch contains LSTM units, and a cross-branch attention gate is added during the cell state update of the LSTM; this gate calculates the current LSTM hidden state and the fluid branch context. The correlation, and based on this, decide from The model also includes a micro-physics solver. After the network outputs the original temperature-time series, it is not directly passed on, but is input into a differentiable simplified heat transfer model for verification. The solver calculates the theoretical thermal stress distribution under this temperature curve. If it exceeds the safety threshold, it generates a gradient signal to correct the temperature curve in reverse. Second branch network input: ;in, For the context from the branch of thermodynamics, The quality potential gradient is used for the quality branch; the network structure is based on a residual decision network; the output is an adaptive fuel gas ratio to environmental conditions. and rotational speed ; The third branch is the quality prediction branch, and the network input is... The output includes two parallel output heads; output head A is the state prediction, and the output is the standard prediction quality vector. The output head B is a gradient prediction, which outputs an approximation of a Jacobian matrix J, where J[i, j] represents the marginal effect of a small change in the j-th process parameter on the i-th quality index; this is achieved by designing dual forward propagation in the network and combining it with automatic differentiation techniques. The physical information exchange and collaboration network receives intermediate layer features or preliminary outputs from three branches, performs calculations, and then feeds the results back to the control parameter branch. First, it performs heat-fluid coupling calculations, using a built-in simplified, differentiable physical calculation unit to calculate a theoretically required heat flux intensity value based on the temperature curves of the thermodynamic branch. Simultaneously, based on the gas ratio in the fluid dynamics branch, a theoretical flame temperature value (Flametemp) is calculated; the target gap signal is calculated, and the output of the mass prediction branch is used to determine the target gap signal. With the highest quality target quality difference between The obtained physical information exchange and collaborative network includes generating a consistency signal, calculating the degree of matching between the theoretically required heat flux intensity value and the theoretical flame temperature value, i.e., the thermodynamic matching degree; and applying the thermodynamic matching degree... and quality difference As additional contextual features, they are sent back to the intermediate layer of the thermodynamics branch and the fluid mechanics branch, respectively; Physical constraint output layer, input adjusted The processing includes numerical range constraints and temporal relationship constraints; The numerical range constraint is: the output is strictly limited to the process safety window using an activation function; The temporal relationship constraint ensures the physical rationality of the temperature curve through constructive functions; Output the final feasible set of circular nozzle process parameters.

2. The adaptive adjustment system for process parameters in ampoule bottle making according to claim 1, characterized in that, The cross-sectional roughness is obtained by acquiring the cross-sectional profile using a white light confocal displacement sensor, which generates multiple profile lines of the entire cross-section, forming a cross-sectional point cloud. The cross-sectional roughness is obtained by calculating the arithmetic mean deviation of all point clouds from the fitted plane.

3. The adaptive adjustment system for process parameters in ampoule bottle making according to claim 2, characterized in that, The second branch network is based on a residual decision network and includes a basic decision path and a correction path. The basic decision path directly generates the initial action from the cross-sectional feature vector, while the correction path integrates the context of the thermodynamic branch and the mass potential gradient to generate a correction quantity. The output is the sum of the initial action and the correction quantity.

4. The adaptive adjustment system for process parameters in ampoule bottle making according to claim 3, characterized in that, The quality potential gradient in the third branch network is calculated as follows: Where ΔQ is the target mass With prediction quality The difference, where J is the Jacobian matrix.

5. The adaptive adjustment system for process parameters in ampoule bottle making according to claim 1, characterized in that, The loss function expression for the hierarchical multi-branch physical information neural network is as follows: in, The main prediction loss is the mean square error between the model's final output and the actual optimal process parameters. Loss due to physical laws; This results in a loss of coordination and consistency. To aid in loss, we characterize the difference between the output of the quality prediction branch and the actual quality.

Images