A method for injection molding and filling a plastic terminal block

Abstract

This invention relates to the field of industrial control technology, and more specifically, to a method for controlling the injection filling of plastic junction boxes. Within a narrow window from 85%-95% filling to the initial holding pressure stage, a deep Q-network model adaptively determines the frequency, power, start-up timing, holding pressure duration, and waveform parameters of ultrasonic vibration based on real-time sensing data such as mold cavity pressure and temperature. High-frequency shearing promotes the diffusion and entanglement of molecular chains at the weld line interface, while the buffering effect of cavitation microbubbles suppresses shrinkage marks. The system employs digital twin pre-training and transfer learning, and incorporates action mask security constraints and multi-level interpretability modules. This invention achieves synergistic suppression of two typical defects—weld lines and shrinkage marks—within a single process window, significantly reducing the number of trial moldings and injection pressure, saving raw materials, and shortening the development cycle.

Description

Technical Field

[0001] This invention relates to the field of industrial control technology, and more specifically, to a method for controlling the injection molding filling of a plastic junction box. Background Technology

[0002] Plastic junction boxes, due to their complex internal structure including reinforcing ribs, mounting posts, and thin-walled partitions, exhibit two typical quality defects during injection molding: weld lines and shrinkage marks. Weld lines are weak areas formed at the junction of two material flows due to the drop in temperature at the melt front and the failure of molecular chains to fully entangle. In severe cases, this can cause the product to crack along the weld line under stress. Shrinkage marks, on the other hand, occur during the cooling and shrinkage process in areas with larger wall thicknesses, resulting in surface depressions or internal holes due to insufficient shrinkage.

[0003] The improvement paths for the two types of defects mentioned above in traditional processes are fundamentally contradictory: increasing injection speed or melt temperature helps reduce weld lines, but it exacerbates uneven cooling and increases the risk of shrinkage marks; extending holding time or increasing holding pressure helps reduce shrinkage marks, but it may affect the long-term reliability of the weld line area due to increased residual stress. Existing technologies typically adopt a "divide and conquer" optimization strategy.

[0004] In recent years, ultrasonic vibration-assisted injection molding technology has attracted widespread attention. Studies have shown that applying ultrasonic vibration immediately after resin filling can increase the weight of optical lenses, and the shrinkage during the holding pressure stage is suppressed by the oscillating flow generated by ultrasonic vibration.

[0005] However, existing technologies still have the following technical shortcomings: First, ultrasonic vibration parameters mostly use fixed values ​​or simple logic control, lacking the ability to perceive and adaptively adjust the state of the mold cavity in real time. Second, most solutions only target a single defect and fail to achieve synergistic suppression of two types of defects within the same process window. Third, due to the lack of an intelligent closed-loop control architecture, it is difficult to adapt to dynamic changes in production conditions such as raw material batch fluctuations and changes in ambient temperature.

[0006] Therefore, developing an adaptive control method and system that can utilize deep reinforcement learning to autonomously decide on ultrasonic vibration strategies, simultaneously suppressing both weld lines and shrinkage marks within the same process window, and systematically solving the interpretability problem of deep Q-networks, has significant engineering application value and broad market prospects. Summary of the Invention

[0007] This invention aims to provide a method and system for controlling the injection molding filling of plastic junction boxes, addressing the shortcomings of existing technologies. These technologies often employ fixed values ​​or simple logic control for ultrasonic vibration parameters, lacking real-time sensing and adaptive adjustment capabilities for mold cavity states. Most solutions only address single defects, failing to achieve synergistic suppression of two types of defects within the same process window. Furthermore, the lack of an intelligent closed-loop control architecture makes it difficult to adapt to dynamic changes in production conditions such as raw material batch fluctuations and ambient temperature variations. To solve these technical problems, the first aspect of this invention provides a method for controlling the injection molding filling of plastic junction boxes, comprising the following specific steps: Arrange cavity pressure sensors and temperature sensor arrays at key locations in the injection mold to construct a real-time sensing network for the melt filling state; A deep Q-network model is established. The input of the deep Q-network model is the cavity pressure and temperature data collected by the real-time sensing network, and the output is the ultrasonic vibration initiation timing, duration, vibration frequency, and vibration amplitude. During the injection molding filling stage, the cavity filling amount is detected. When it reaches a preset threshold, the deep Q-network model is triggered to perform inference and output the current optimal ultrasonic vibration control strategy. According to the ultrasonic vibration control strategy, within the time window from when the filling amount reaches the threshold to when the pressure holding begins, high-frequency shear vibration is applied to the melt through the ultrasonic transducer preset at the mold. During the pressure holding stage, ultrasonic vibration is continued to be applied to form a uniformly distributed cluster of microbubbles inside the melt. The quality data of the junction box product after vibration treatment is obtained, and the quality data is used as a reward feedback signal to input into the experience playback pool to iteratively update the deep Q-network model.

[0008] Preferably, the key locations of the injection mold include the melt confluence, the thick-walled area, and the gate.

[0009] Preferably, the ultrasonic vibration control strategy includes start-up timing, frequency, power amplitude, and waveform modulation parameters.

[0010] Preferably, the reward function of the deep Q-network model includes both the weld line strength evaluation index and the shrinkage depth evaluation index, enabling the deep Q-network model to learn a collaborative optimization strategy that can improve both weld lines and shrinkage defects during the training process. The physical effects of ultrasonic vibration include promoting the interdiffusion and entanglement of molecular chains at the weld line interface, and utilizing the cavitation effect to form microbubble groups in the melt, which play a buffering and filling role during cooling and contraction. The vibration frequency range for promoting the interdiffusion and entanglement of molecular chains is 30–100 kHz, and the vibration frequency range for the cavitation effect is 20–60 kHz.

[0011] Preferably, the deep Q-network model adopts a three-layer fully connected neural network structure. The input layer receives real-time signals collected by the cavity pressure sensor and temperature sensor, the hidden layer contains 64 to 256 neurons, and the output layer outputs the continuous or discrete values ​​of the ultrasonic vibration control strategy.

[0012] Preferably, the method further includes a digital twin pre-training step: establishing a digital twin model of the junction box based on injection mold flow analysis software, pre-training a deep Q-network model in a virtual environment to obtain an initial control strategy; then transferring the initial control strategy to a real injection molding machine through transfer learning, and fine-tuning it according to actual production data.

[0013] Preferably, the waveform modulation parameters are selected from one or more combinations of continuous wave mode, pulse wave mode, or frequency sweep mode; When an abnormal pressure attenuation is detected at the melt confluence, the deep Q-network model adjusts the waveform from a continuous wave to a pulse wave mode. When the temperature drop rate in the thick-walled region exceeds a preset threshold, the deep Q-network model adjusts the vibration frequency to the optimal frequency range for the cavitation effect.

[0014] A second aspect of the present invention provides a plastic junction box injection filling control system, which uses the above-described method for control, including: Sensing layer: Cavity pressure sensor array and temperature sensor array deployed at key locations of the injection mold, as well as melt front arrival sensor array, are used to collect cavity status data in real time during the injection filling process. Key locations include melt confluence, thick-walled areas and gates. Execution layer: includes at least one ultrasonic transducer, the end face of which forms part of the inner wall of the mold cavity; the ultrasonic transducer is connected to an ultrasonic generator via a signal line and is used to apply high-frequency shear vibration to the melt in the mold cavity according to control commands; Control layer: includes a central controller and a pre-trained deep Q network model; the central controller receives real-time cavity pressure and temperature data collected by the perception layer, calls the deep Q network model for inference, and outputs the frequency, amplitude and timing instructions of ultrasonic vibration to the execution layer; The self-evolutionary layer includes an experience replay storage module and a transfer learning module. The experience replay storage module is used to store mold cavity state data, vibration strategies, and product quality data during the injection molding production process. The transfer learning module is used to periodically retrain the deep Q-network model offline or update it online based on historical production data.

[0015] Preferably, the number of ultrasonic transducers is 1 to 4, each corresponding to a different vibration region of the mold cavity; The output power of the ultrasonic generator is adjustable from 0 to 1000W, and the vibration frequency range is 20 to 100kHz.

[0016] Preferably, the sensor for the arrival of the melt front is a fiber optic detection sensor or a piezoelectric acoustic wave sensor. The sensing layer also includes a cavity displacement sensor, which is used to monitor the cavity deformation in the thick-walled area in real time to calculate the shrinkage depth; the measurement accuracy of the cavity displacement sensor is not less than ±0.5µm.

[0017] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0018] A fourth aspect of the present invention provides a computer device including a memory and a processor, the memory storing a computer program: when the processor executes the computer program, it implements the steps of the above-described plastic junction box injection filling control method.

[0019] Compared with the prior art, the present invention has the following beneficial effects: This invention achieves synergistic suppression of two typical defects in plastic junction boxes—weld marks and shrinkage marks—in the same process step by fusing deep reinforcement learning with narrow-window ultrasonic vibration from the end of filling to the beginning of holding pressure. Experimental results show that compared with existing technologies, the length, width, and depth of weld marks are reduced by 52.7%, 41.4%, and 52.4%, respectively, while the elastic modulus and hardness are increased by 12.5% ​​and 17.2%, respectively; the shrinkage mark depth is reduced by 62.0%, the shrinkage rate is reduced by 68.8%, the peak injection pressure is reduced by 57.3%, and raw materials are saved by 13%; the improvement effects on the two types of defects show a significant positive coupling effect. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the drawings without creative effort. In the drawings: Figure 1 This is a flowchart of a plastic junction box injection filling control method according to an embodiment of the present invention.

[0021] Figure 2 This is a system diagram of a plastic junction box injection filling control system according to an embodiment of the present invention. Detailed Implementation

[0022] Hereinafter, exemplary embodiments according to the present invention will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments of the present invention. It should be understood that the present invention is not limited to the exemplary embodiments described herein.

[0023] As stated in the background section above, the present invention aims to adaptively optimize the ultrasonic vibration strategy in the transition phase between filling and holding pressure in a single process, while suppressing both weld lines and shrinkage marks in plastic junction box injection molded parts.

[0024] The core of this invention lies in constructing a four-layer closed-loop intelligent control architecture of perception, decision-making, execution, and evolution. Unlike traditional fixed-parameter ultrasonic-assisted injection molding, this system uses a deep reinforcement learning agent to dynamically decide the timing, duration, frequency mode, and amplitude of ultrasonic vibration based on the real-time mold cavity state. Within the "golden time window" from the end of filling to the beginning of holding pressure, a single set of control commands simultaneously achieves two major objectives: 1. Promoting the full diffusion and entanglement of molecular chains in the weld line region; 2. Alleviating volume shrinkage in thick-walled regions through the uniform distribution of cavitation microbubbles.

[0025] The system architecture consists of the following four parts: Sensing layer: Mold cavity pressure sensors and temperature sensor arrays are arranged at key locations in the mold, such as melt confluence, thick-walled areas, and gates, and a melt front arrival sensor is added to build a real-time multi-dimensional sensing network for the internal state of the mold cavity.

[0026] Decision layer: The intelligent agent is trained using the Deep Q-Network (DQN) algorithm. The agent's state space consists of the current sensor data stream and the product's 3D model features, while its action space consists of the ultrasonic vibration start / stop signal, vibration frequency, vibration amplitude, and vibration waveform modulation parameters.

[0027] Execution layer: The end face of the ultrasonic transducer is directly incorporated into the inner wall of the mold cavity. Driven by the ultrasonic generator, it applies precise high-frequency shear vibration to the polymer melt in the mold cavity according to the output command of the Agent.

[0028] Self-evolution layer: Based on historical production data, deep reinforcement learning models are retrained offline or transferred online to achieve continuous optimization of control strategies.

[0029] The ultrasonic vibration parameters involved in this invention include five main categories: vibration frequency, power amplitude, initiation timing, duration of action, and waveform modulation parameters. These parameters exhibit complex coupling relationships. Ultrasonic-assisted injection molding can reduce melt viscosity by approximately 38%, improve mechanical properties by approximately 10.21%, and increase crystallinity by approximately 17.1%. The application stage and duration of ultrasonic vibration have a significant impact on melt filling behavior and product performance; applying it during the filling stage produces drastically different physical effects compared to applying it after filling.

[0030] 1. Selection of ultrasonic frequency: Ultrasonic frequency directly affects two types of physical effects: (1) Molecular chain diffusion promotion effect, frequency range: 30~100kHz; within this frequency range, high frequency shear stress can cause molecular chains at the melt interface to reorient and entangle with each other, significantly improving the mechanical properties of the weld line area.

[0031] (2) Cavitation microbubble effect, frequency range: 20-60kHz; within this frequency range, ultrasonic vibration can induce the formation and collapse of cavitation bubbles in the melt. Lower frequencies of 20-40kHz are more likely to produce cavitation, and 30-40kHz is the optimal frequency range for producing cavitation.

[0032] The frequency output of this invention is not a single fixed value, but rather dynamically switches or mixes between two frequency bands according to the current process state and defect suppression requirements. When the system detects that melt fusion is about to occur, i.e., weld lines are about to form, the DQN Agent preferentially outputs mid-frequency vibrations of 40–60 kHz to take into account both molecular chain diffusion and slight cavitation; when the thick-walled region begins to cool and shrink, i.e., shrinkage marks begin to form, the Agent switches to low-frequency vibrations of 20–40 kHz to enhance the formation efficiency of cavitation microbubbles.

[0033] 2. Selection of power amplitude: The ultrasonic power amplitude is controlled by the output power of the transducer and can be adjusted within the range of 0 to 1000W. The relationship between the power amplitude and the vibration effect is non-linear. Low power range, 0-300W; mainly generates heat through molecular chain vibration and viscoelasticity, with local temperature increase of about 2-8°C in the melt, suitable for preventing premature cooling of the melt in the initial stage of filling. Medium power range, 300-600W; molecular chain diffusion is accelerated, cavitation bubbles begin to form, suitable for improving weld lines and pre-generating microbubbles from the end of filling to the beginning of holding pressure; High power range, 600-1000W; strong cavitation effect, generating a large number of microbubbles, suitable for suppressing shrinkage marks in thick-walled regions.

[0034] 3. Selection of waveform modulation The ultrasonic generator of this invention supports the selection and dynamic switching of three basic waveforms: (1) Continuous wave mode: The ultrasonic waves are continuously output with high energy density, which is suitable for scenarios that require continuous shearing and heating, such as the filling stage when the melt temperature is low. Continuous waves can provide a stable molecular chain diffusion environment, but they have high requirements for the heat load of the transducer.

[0035] (2) Pulse wave mode: The ultrasonic power is output intermittently at a certain duty cycle of 10% to 90%, with low energy density but small heat accumulation. Pulse waves are suitable for temperature-sensitive material systems or when abnormal pressure attenuation is detected at the junction.

[0036] (3) Frequency sweep mode: The vibration frequency is cyclically scanned within a preset frequency band, with a scanning period of 1 to 5 seconds, so that the melt is subjected to alternating shearing action at different frequencies. The frequency sweep mode is particularly suitable for situations where there are large batch fluctuations in materials or where it is necessary to simultaneously consider molecular chain diffusion and cavitation effects.

[0037] 4. Timing of application: The decision on when to initiate ultrasonic vibration is based on the filling volume detection results. This invention limits the initiation window to 85%–95% of the filling volume, and this selection is based on the following physical principles: Premature initiation (<85%): The melt has not yet merged or the thick-walled region has not yet formed. Most of the vibration energy is consumed in the melt of the main channel, and the improvement effect on the two types of defects is not significant. Late start-up (>95%): The melt basically fills the cavity, the weld line interface has partially solidified, and the molecular chain diffusion ability has decreased significantly; at the same time, the cooling of the thick-walled area has reached a considerable level, making it difficult for cavitation microbubbles to form.

[0038] During the pressure holding phase, ultrasonic vibration continues to be applied for 1 to 2 seconds. Its mechanism of action changes from molecular chain diffusion to microbubble group buffering and filling. Under the pressure holding pressure, the cavitation microbubbles that have been formed are partially compressed but do not completely disappear, forming a microporous elastic skeleton structure. When the thick-walled area cools and shrinks, it absorbs part of the volume shrinkage through compression deformation, thereby effectively inhibiting the formation of shrinkage marks.

[0039] The filling amount test result can be calculated based on the integral of the mold cavity pressure or the screw position. Specifically: The decision-making basis for the ultrasonic vibration initiation timing is the accurate detection of the real-time filling volume. This embodiment uses two methods to independently calculate the filling volume and improves detection accuracy and robustness through data fusion.

[0040] Method 1: Filling amount estimation based on screw position sensor The injection screw of an injection molding machine moves forward under the drive of the injection cylinder, pushing the molten metal into the mold cavity. There is a definite correspondence between the screw's displacement stroke and the volume of molten metal injected into the cavity. Let the screw diameter be... The screw stroke is The theoretical injection volume for:

[0041] Considering the compressibility of the melt under high temperature and high pressure, the compressibility coefficient In addition to pressure losses in the injection molding machine nozzle and runner, the actual volume of melt entering the mold cavity. Corrections are needed. Assume the current mold cavity pressure is... Atmospheric pressure is The volume compression correction factor is:

[0042] The retention volume of melt in the flow channel This is a fixed value, obtainable through mold flow analysis or experimental calibration. Therefore, the actual volume of melt entering the mold cavity is:

[0043] Let the total volume of the cavity be The current filling percentage can be accurately measured using a 3D model of the mold. for:

[0044] Screw stroke Real-time measurement is achieved using a magnetostrictive displacement sensor mounted on the injection cylinder, with a measurement accuracy of ±0.01 mm and a sampling period of 10 ms. When the preset threshold is reached, ultrasonic activation is triggered.

[0045] Method 2: Calculation of filling amount based on cavity pressure integral When the melt flows through the location of the pressure sensor in the mold cavity, the sensor outputs a pressure signal. It changes over time. During the filling phase, there is a monotonic relationship between the integral area of ​​the pressure curve and the filled volume. Let's assume that from the start of filling... up to the current moment The pressure integral is:

[0046] Through offline calibration, under the same injection molding process parameters, the pressure integral values ​​corresponding to different filling amounts were measured, and a mapping function was established. , so that:

[0047] Calibration method: Set a transparent window on the mold or use a short-shot test. Forcefully stop injection when the filling amount reaches the predetermined value, remove the incompletely filled product, weigh it, and establish a nonlinear mapping table between pressure integral and filling amount. In real-time control, this is quickly obtained through table lookup and linear interpolation. .

[0048] Under normal circumstances, the calculation using the screw position method The primary value is used; when an abnormality is detected in the screw position sensor or abnormal pressure fluctuations, the system automatically switches to the pressure integration method as a backup. If the calculation results from the two methods deviate by more than ±3%, the system issues a sensor calibration alarm.

[0049] 5. Selection of duration of action The duration of ultrasonic vibration is not independent of the initiation timing parameter, but is determined by both the initiation and termination timing. This invention decomposes the ultrasonic vibration process into two stages, the duration of which is controlled by different factors and is coupled with each other.

[0050] Vibration duration during filling period It is not an independent parameter directly output by the Agent, but is determined by both the startup timing and the population completion time:

[0051] in It is the absolute moment when the filling is complete. A signal is triggered when the fill rate reaches 100%. Since the fill rate is not constant during injection molding, the agent cannot directly predict it. However, this problem can be solved in this invention through the following steps: The Agent is incorrect. Instead of direct control, This is considered a consequence of the launch timing decision. When the agent chooses an earlier launch timing, for example... = 85%, Automatic startup time is longer; when a later startup time is selected, such as... =93%, Automatically shorter.

[0052] Vibration duration during pressure holding period The decision is made independently by the Agent. It is one of the core parameters that the Agent needs to make independent decisions. Its output range is a continuous value of 0.5 to 3.0 seconds, with a step accuracy of 0.1 seconds.

[0053] For different materials, a benchmark value can be set based on the material properties: ABS / PC / PC-ABS: Baseline = 1.5–2.0 seconds PP / PE (crystalline plastic): Baseline = 1.0 to 1.5 seconds PA6 / PA66 (high shrinkage): Baseline = 2.0 to 2.5 seconds PMMA / PS (Amorphous Brittleness): Baseline = 1.0–1.8 seconds The sensing layer monitors the temperature of the thick-walled region in real time. and its rate of decline .when When this occurs, it indicates that the melt is rapidly solidifying and the cavitation window is about to close; the agent will... Reduce the current baseline value by 0.2 to 0.5 seconds; when When this indicates good melt flowability, the Agent will... Increase by 0.3 to 0.6 seconds.

[0054] During the pressure holding phase, the primary role of ultrasonic vibration shifts from promoting molecular chain diffusion to microbubble buffering and filling. This is achieved when the mold cavity pressure at the weld line converges... attenuation rate When the vibration level drops to near zero and persists for more than 0.3 seconds, it indicates that the area has essentially solidified, and continued vibration is detrimental to the product. At this point, the Agent issues an early shutdown command, regardless of... Has the preset value been reached?

[0055] By analyzing the high-frequency signals from the cavity displacement sensor, the microbubble resonance characteristics are analyzed. When the average size of the microbubble is detected to be >50 micrometers, the agent immediately shuts off the ultrasound to prevent the microbubbles from merging and forming macroscopic pores.

[0056] Total duration of ultrasound treatment The optimal strategy learned by the agent during training is to minimize the impact of both types of defects while ensuring the improvement of both types of defects. This is to save energy, reduce heat accumulation, and shorten cycle time.

[0057] The collaborative optimization logic is as follows: When weld line risk is dominant: the agent should start earlier. =86%-88%, making The duration is relatively long, approximately 200–400 ms, which fully promotes molecular chain diffusion; at the same time, it is appropriately shortened. 1.0 to 1.5 seconds, to avoid excessive total duration leading to heat accumulation.

[0058] When shrinkage risk is dominant: the agent chooses to start later. = 92%-94%, Shorter, approximately 100–200 ms; while extended 2.0 to 2.5 seconds to ensure sufficient formation of cavitation microbubbles and maintenance of the buffer structure.

[0059] When both types of risk are high: the agent chooses a moderate launch time. = 89%-91%, and Distribute the time evenly to each of the segments, approximately 1.0 to 1.5 seconds, for a total duration of [duration missing]. Keep it between 2.0 and 2.5 seconds.

[0060] This embodiment employs a three-level closed-loop control architecture to achieve precise execution of the ultrasonic vibration strategy: First-stage closed loop, resonant frequency automatically tracked in real time: During injection molding, the resonant frequency of ultrasonic transducers drifts due to factors such as changes in load temperature and fluctuations in melt pressure. Without compensation, this leads to a sharp decrease in energy transmission efficiency. This invention integrates an automatic resonant frequency tracking module into the ultrasonic generator, employing a combination of phase-locked loop (PLL) and fuzzy proportional-integral control to track the frequency. This method establishes the transducer's frequency characteristic equation based on the Mason equivalent circuit, selects the series resonant point as the operating frequency of the ultrasonic processing system, and continuously adjusts the drive signal frequency by acquiring the transducer's operating current and voltage phase difference signals in real time, aiming to bring the phase difference close to zero. The tracking error accuracy is controlled within 0.82%.

[0061] Second-stage closed-loop, amplitude adaptive constant control: The ultrasonic generator also features a constant power amplitude control loop. This loop monitors the transducer's output power and ultrasonic energy in real time, and adjusts the amplitude of the drive voltage using a fuzzy PI control algorithm to ensure that the output power remains stable near the target value set by the Agent. When an output power deviation exceeds a set threshold, such as 5%, the amplitude adjustment loop responds quickly, causing the power output to rapidly return to the target value.

[0062] The third-level closed loop involves dynamic policy adjustment driven by DQN: All output parameters of the ultrasonic generator, including start / stop signals, frequency, power amplitude, and waveform mode, are dynamically determined by a deep Q-network model in the central controller based on the real-time cavity status. This decision layer receives sensor data from the sensing layer at 10ms intervals, including cavity pressure, temperature, and melt front arrival signals, and outputs optimal control commands through neural network inference.

[0063] The three-loop collaborative mechanism is as follows: the first-level closed loop ensures that the energy transmission efficiency is maximized, the second-level closed loop ensures that the output energy is consistent with the set value, and the third-level closed loop adjusts the set value in real time according to the current process status, thus forming a fully closed-loop intelligent control system.

[0064] The DQN model of this invention adopts a three-layer fully connected neural network structure. The input layer is a 16-dimensional sensor data vector, the hidden layers are two-layer structures with 128 and 64 neurons respectively, and the output layer is a 4-dimensional action space. Each layer of the network employs explicit physical semantics. The input layer includes: 3 cavity pressure sensors, 3 temperature sensors, 3 melt front arrival signals, 1 screw stroke signal, 1 fill percentage, 1 pressure decay rate, 1 temperature drop rate, 2 residual stress indicators, and 1 batch code. Each input is normalized and has a dimensionless range of [-1, 1] or [0, 1].

[0065] The hidden layers consist of: a first hidden layer with 128 neurons, used to capture the nonlinear coupling relationship between sensor data; and a second hidden layer with 64 neurons, used to extract higher-level features related to ultrasonic vibration decision-making.

[0066] The output layer includes four control parameters for output ultrasonic vibration: frequency, power amplitude, start-up timing, and waveform modulation parameters.

[0067] Experience replay pool: with a capacity of 10,000, used to store state transition data during the training process.

[0068] This embodiment constructs a four-layer granularity DQN decision process visualization and monitoring interface: Training Cycle Layer: Displays the convergence curve of the cumulative reward throughout the entire training process. Reference convergence value: The cumulative reward is stable in the range of >2.0, with a fluctuation range of <±0.3.

[0069] Process Cycle Layer: Displays the total reward for each injection molding process cycle and the contribution breakdown of each sub-objective, weld line strength factor, and shrinkage mark inhibition factor.

[0070] Stage layer: Each process cycle is decomposed into the filling stage (0-85%), the transition stage (85%-95% to the start of holding pressure) and the holding pressure stage, respectively displaying the action sequence and state change trajectory within each stage.

[0071] Step-level: Displays sensor data for each step (control cycle 10ms), Q-value distribution of DQN output, and the final selected action command at millisecond time resolution.

[0072] The visualization system uses t-SNE dimensionality reduction technology to project the high-dimensional state space onto a two-dimensional plane and cluster similar states for display.

[0073] This invention introduces a self-attention mechanism module after the first hidden layer of DQN. By learning the attention weight matrix, it automatically identifies the most critical sensor input features in the current decision. The attention score is calculated based on the similarity between the query vector and the key vector. The normalized attention score is presented in real-time as a heatmap on the monitoring interface; a higher score indicates a greater impact of the sensor input on the current decision.

[0074] This invention employs two model interpretability algorithms, SHAP and LIME, to perform offline feature attribution analysis on DQN decisions: SHAP analysis: Based on the Shapley value in game theory, it calculates the marginal contribution of each input feature to the final decision and outputs a feature contribution ranking chart.

[0075] LIME analysis: For a specific decision, it fits an interpretable linear model within a local region. For example, LIME analysis might output: Under the current conditions of cavity pressure 42MPa, temperature 235℃, and fill level 92%, the main reasons for choosing 450W power are, in order, temperature lower than the target value, pressure rise rate slower, and fill level close to the threshold.

[0076] This invention employs an offline decision tree fitting method to "distill" complex DQN decision strategies into a readable IF-THEN rule set. It collects a large amount of (state, action) data pairs generated by the agent during actual production processes, constructs a CART decision tree model, and uses the state space of the deep Q-network as input and the action chosen by the agent or the Q-value distribution in that state as output for supervised learning training.

[0077] This invention provides a reward breakdown visualization function, displaying the composition of accumulated rewards in real time on the monitoring interface. The reward value is calculated separately for four sub-items and displayed in real time in the form of a stacked bar chart: Weld line strength bonus: Based on the pressure attenuation characteristics of the joint area detected by the mold cavity pressure sensor, the value range is [0, 1].

[0078] Sink mark suppression reward: Based on the cavity displacement sensor to detect the cavity deformation in the thick-walled region, the value range is [0, 1].

[0079] Residual stress penalty: estimated based on the cavity pressure decay rate at the end of the holding pressure, with a value range of [0, 0.5].

[0080] Cycle time penalty: The deviation of the current process cycle time from the target cycle time, with a value range of [0, 0.3].

[0081] To facilitate a further understanding of the present invention, specific examples are used below to provide a detailed description of the invention: Example

[0082] This embodiment provides a collaborative control method for injection molding of plastic junction boxes based on deep reinforcement learning, which is used for injection molding of junction boxes made of ABS material. The junction box has external dimensions of 120mm×80mm×40mm, an average wall thickness of 2.5mm, and a maximum wall thickness of 5mm.

[0083] Step 1: Build the system hardware platform A junction box mold equipped with an ultrasonic vibration device is installed on an 80-ton reciprocating screw injection molding machine. The end face of the ultrasonic transducer forms part of the inner wall of the mold cavity. The transducer is connected to an ultrasonic generator via a signal line. The output power of the ultrasonic generator is adjustable from 0 to 1000W, and the frequency range is 20 to 100kHz. The following sensors are placed at key locations on the mold: (1) One cavity pressure sensor is arranged at the gate, the melt junction and the thick-walled area, with a range of 0 to 200 MPa and an accuracy of ±0.5%FS; (2) Arrange a K-type thermocouple temperature sensor in the same position as above, with a measurement range of 0 to 400℃ and an accuracy of ±1℃; (3) Arrange three melt fronts to reach the sensor on the critical flow path of the cavity; (4) Arrange cavity displacement sensors in the thick-walled area with an accuracy of ±0.5µm.

[0084] Step 2: Construct a deep reinforcement learning model A deep Q-network (DQN) was established with the following structure: the input layer is a 16-dimensional vector containing real-time data such as cavity pressure, temperature, and melt front arrival time; the hidden layers are two fully connected layers, containing 128 and 64 neurons respectively, with ReLU as the activation function; the output layer is a 4-dimensional vector, corresponding to continuous values ​​of ultrasonic frequency from 20 to 60 kHz, continuous values ​​of ultrasonic power from 0 to 600 W, relative values ​​of fill volume from 85% to 95% at startup, and waveform modulation parameters of continuous wave / pulse wave / sweep mode.

[0085] The reward function R is set as follows:

[0086] In the above formula, This is the weld line strength factor. It is a shrinkage inhibition factor. For residual stress, The loop time; These are the corresponding weighting coefficients, with values ​​of 0.4, 0.3, 0.2, and 0.1, respectively. The weld line strength factor is indirectly evaluated by detecting the pressure attenuation characteristics of the fusion region using a mold cavity pressure sensor; the shrinkage mark suppression factor is monitored in real time by a mold cavity displacement sensor in the thick-walled region; and the residual stress is estimated based on the mold cavity pressure attenuation rate at the end of the holding pressure test. The cycle time is the normalized deviation of the current process cycle time from the target cycle time, with a difference of 15 seconds.

[0087] Step 3: Digital Twin Pre-training A digital twin model of the junction box was built using Moldflow software, with ABS as the material. The digital twin model was then connected to a DQN agent. The operating parameters of the digital twin model included: injection time 1.2 seconds, melt temperature 230℃, mold temperature 50℃, holding pressure 60MPa, and holding time 3 seconds. Pre-training was performed in a virtual environment with 2000 episodes, each episode executing 100 control steps, for a total of 200,000 steps. The experience replay pool capacity was set to 10,000, the mini-batch size to 64, the discount factor γ = 0.99, the learning rate α = 0.001, and the target network update frequency to be synchronized every 100 steps.

[0088] Step 4: Filler volume detection and ultrasonic initiation decision The injection molding cycle is initiated, and the sensing layer collects multi-point pressure and temperature data of the mold cavity at 10ms intervals. When the fill volume reaches a preset threshold of 92% ± 1%, the DQN Agent outputs the optimal ultrasonic vibration strategy based on the current mold cavity state. In this embodiment, the optimal strategy output by the Agent under steady-state conditions is: [Initiation timing information would be inserted here]. The fill power is 92%, the frequency is 45kHz, the power is 400W, and the waveform is a continuous wave. =1.5 seconds. The actual filling rate can be calculated. It is 100ms. It takes 1.6 seconds.

[0089] Step 5: Apply ultrasonic vibration and adjust in real time. During the time window from the end of filling to the beginning of holding pressure, the ultrasonic transducer is driven to perform high-frequency shear vibration according to the ultrasonic strategy output by the Agent. During vibration, the Agent continuously receives feedback signals from the cavity sensor. When abnormal changes in melt flow characteristics are detected, the vibration parameters are adjusted. Specifically, if the temperature drop rate in the thick-walled region exceeds 5°C / second, the Agent automatically increases the ultrasonic power from 400W to 500W to enhance the cavitation effect; if the pressure attenuation in the confluence region is abnormal, the waveform is adjusted from a continuous wave to a pulsed wave.

[0090] Step 6: End of pressure holding and product inspection After the pressure holding stage, the ultrasonic vibration is turned off, and the cooling stage begins. After the product is demolded, the contour dimensions and shrinkage depth of the weld line area are measured using a non-contact three-dimensional optical scanning system.

[0091] Step 7: Model Update and Policy Evolution Defect detection data from finished products is used as a reward feedback signal and input into the experience replay pool. After every 100 production batches, a mini-batch is randomly sampled from the experience replay pool for offline retraining. When a quality indicator (such as weld line strength factor) is detected to have decreased continuously by more than 5%, online transfer learning updates are initiated.

[0092] Example 2 The difference between this embodiment and Embodiment 1 is that the timing of the ultrasonic vibration is adjusted to start at 88% fill volume, hold for 1 second, and then stop. The waveform is also adjusted to a sweep frequency mode, cyclically scanning from 30 to 60 kHz with a period of 2 seconds. This embodiment is used to verify the effect of different ultrasonic strategies on the synergistic suppression of weld lines and contraction marks.

[0093] Example 3 The difference between this embodiment and Embodiment 1 is that polypropylene (PP) is used as the injection molding material to verify the adaptability of the present invention to different material systems. In this embodiment, the recommended process parameters for PP material are: melt temperature 220°C, mold temperature 40°C, and injection speed slightly lower than that for ABS.

[0094] Example Performance Data Under the process conditions of Example 1, the product was tested after injection molding, and the following data were obtained: (1) Weld line suppression effect:

[0095] (2) Sinking mark / pore suppression effect:

[0096] Control experiment: Group A used only the "92% start-up at the end of the filling stage + continuous wave ultrasound" method during the filling phase; Group B used only the "multi-objective reinforcement learning parameter optimization" method without ultrasonic vibration; Group C used the complete scheme of this invention. The results are as follows:

[0097] During the operation of Example 1, the interpretability monitoring module of this invention was activated. The operator observed through the visualization interface that when the viscosity of a batch of ABS raw material was low, the attention heatmap showed that the attention scores for the two features—mold cavity pressure and temperature drop rate—jumped from 0.2 to 0.7. Simultaneously, SHAP analysis showed that these two features contributed the most to frequency selection. Based on this, the operator judged that the agent had correctly adapted to the raw material change, and the operator's trust in the system significantly increased. The action masking mechanism successfully masked the delayed start command when an abnormal sensor signal fill volume false alarm reached 96%, avoiding quality defects caused by premature ultrasonic shut-off.

[0098] The above embodiments and experimental data fully verify the following technical effects of the present invention: (1) By combining narrow-window ultrasonic vibration from the end of filling to the beginning of holding pressure with DRL adaptive parameter optimization, the synergistic suppression of two types of defects, weld lines and shrinkage marks, is achieved in the same process step; (2) The mechanical performance improvement of weld lines and the suppression effect of shrinkage marks are significantly better than the existing technical solutions, and there is an obvious positive coupling effect between the improvement effects of the two types of defects; (3) The DRL adaptive algorithm has excellent environmental adaptability and can cope with the fluctuations of raw material batches and ambient temperature, and continuously maintain the consistency of product quality; (4) Digital twin pre-training significantly reduces the number of physical mold trials and greatly shortens the process development cycle.

[0099] Example 4 This embodiment provides a plastic junction box injection filling control system, which uses the method in Embodiment 1 for control, including: Sensing layer: Cavity pressure sensor array and temperature sensor array deployed at key locations of the injection mold, as well as melt front arrival sensor array, are used to collect cavity status data in real time during the injection filling process. Key locations include melt confluence, thick-walled areas and gates. Execution layer: includes at least one ultrasonic transducer, the end face of which forms part of the inner wall of the mold cavity; the ultrasonic transducer is connected to an ultrasonic generator via a signal line and is used to apply high-frequency shear vibration to the melt in the mold cavity according to control commands; Control layer: includes a central controller and a pre-trained deep Q network model; the central controller receives real-time cavity pressure and temperature data collected by the perception layer, calls the deep Q network model for inference, and outputs the frequency, amplitude and timing instructions of ultrasonic vibration to the execution layer; The self-evolutionary layer includes an experience replay storage module and a transfer learning module. The experience replay storage module is used to store mold cavity state data, vibration strategies, and product quality data during the injection molding production process. The transfer learning module is used to periodically retrain the deep Q-network model offline or update it online based on historical production data.

[0100] The number of ultrasonic transducers is 1 to 4, each corresponding to a different vibration region of the mold cavity; The output power of the ultrasonic generator is adjustable from 0 to 1000W, and the vibration frequency range is 20 to 100kHz.

[0101] The sensor for the arrival of the melt front is either a fiber optic detection sensor or a piezoelectric acoustic wave sensor. The sensing layer also includes a cavity displacement sensor, which is used to monitor the cavity deformation in the thick-walled area in real time to calculate the shrinkage depth; the measurement accuracy of the cavity displacement sensor is not less than ±0.5µm.

[0102] Example 5 According to one aspect of the present invention, a computer program product or computer program is provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and executes the computer instructions, causing the computer device to perform the methods provided in the various alternative implementations described above.

[0103] In another aspect, the present invention also provides a computer-readable medium, which may be included in the electronic device described in the above embodiments; or it may exist independently and not assembled into the electronic device. The computer-readable medium carries one or more programs that, when executed by the electronic device, cause the electronic device to implement the plastic junction box injection filling control method described in the above embodiments.

[0104] It should be noted that although several modules or units of the device for performing actions have been mentioned in the detailed description above, this division is not mandatory. In fact, according to embodiments of the present invention, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.

[0105] Through the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions according to the embodiments of the present invention can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, portable hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, touch terminal, or network device, etc.) to execute the method according to the embodiments of the present invention.

[0106] Those skilled in the art will readily conceive of embodiments of the invention upon consideration of the specification and practice of the methods disclosed herein. The invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein.

[0107] It should be understood that the present invention is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A method for controlling the injection filling of a plastic junction box, characterized in that, The specific steps include the following: Arrange cavity pressure sensors and temperature sensor arrays at key locations in the injection mold to construct a real-time sensing network for the melt filling state; A deep Q-network model is established. The input of the deep Q-network model is the cavity pressure and temperature data collected by the real-time sensing network, and the output is the ultrasonic vibration initiation timing, duration, vibration frequency, and vibration amplitude. During the injection molding filling stage, the cavity filling amount is detected. When it reaches a preset threshold, the deep Q-network model is triggered to perform inference and output the current optimal ultrasonic vibration control strategy. According to the ultrasonic vibration control strategy, within the time window from when the filling amount reaches the threshold to when the pressure holding begins, high-frequency shear vibration is applied to the melt through the ultrasonic transducer preset at the mold. During the pressure holding stage, ultrasonic vibration is continued to be applied to form a uniformly distributed cluster of microbubbles inside the melt. The quality data of the junction box product after vibration treatment is obtained, and the quality data is used as a reward feedback signal to input into the experience playback pool to iteratively update the deep Q-network model.

2. The injection molding filling control method for plastic junction boxes according to claim 1, characterized in that, Key locations in injection molds include the melt confluence, thick-walled areas, and the gate.

3. The injection molding filling control method for plastic junction boxes according to claim 1, characterized in that, Ultrasonic vibration control strategies include start-up timing, frequency, power amplitude, and waveform modulation parameters.

4. The injection molding filling control method for plastic junction boxes according to claim 1, characterized in that, The reward function of the deep Q network model includes both the weld line strength evaluation index and the shrinkage depth evaluation index, enabling the deep Q network model to learn a collaborative optimization strategy that can improve both weld lines and shrinkage defects during training. The physical effects of ultrasonic vibration include promoting the interdiffusion and entanglement of molecular chains at the weld line interface, and utilizing the cavitation effect to form microbubble groups in the melt, which play a buffering and filling role during cooling and contraction. The vibration frequency range for promoting the interdiffusion and entanglement of molecular chains is 30–100 kHz, and the vibration frequency range for the cavitation effect is 20–60 kHz.

5. The injection molding filling control method for plastic junction boxes according to claim 1, characterized in that, The deep Q-network model adopts a three-layer fully connected neural network structure. The input layer receives real-time signals collected by the cavity pressure sensor and temperature sensor, the hidden layer contains 64 to 256 neurons, and the output layer outputs the continuous or discrete values ​​of the ultrasonic vibration control strategy.

6. The injection molding filling control method for plastic junction boxes according to claim 1, characterized in that, It also includes a digital twin pre-training step: establishing a digital twin model of the junction box based on injection mold flow analysis software, pre-training a deep Q-network model in a virtual environment, and obtaining an initial control strategy; The initial control strategy was then transferred to a real injection molding machine through transfer learning and fine-tuned based on actual production data.

7. The injection molding filling control method for plastic junction boxes according to claim 1, characterized in that, The waveform modulation parameters are selected from one or more combinations of continuous wave mode, pulse wave mode, or frequency sweep mode; When an abnormal pressure attenuation is detected at the melt confluence, the deep Q-network model adjusts the waveform from a continuous wave to a pulse wave mode. When the temperature drop rate in the thick-walled region exceeds a preset threshold, the deep Q-network model adjusts the vibration frequency to the optimal frequency range for the cavitation effect.

8. A plastic junction box injection filling control system, controlled by the method described in any one of claims 1-7, characterized in that, include: Sensing layer: Cavity pressure sensor array and temperature sensor array deployed at key locations of the injection mold, as well as melt front arrival sensor array, are used to collect cavity status data in real time during the injection filling process. Key locations include melt confluence, thick-walled areas and gates. Execution layer: includes at least one ultrasonic transducer, the end face of which forms part of the inner wall of the mold cavity; the ultrasonic transducer is connected to an ultrasonic generator via a signal line and is used to apply high-frequency shear vibration to the melt in the mold cavity according to control commands; Control layer: includes a central controller and a pre-trained deep Q network model; the central controller receives real-time cavity pressure and temperature data collected by the perception layer, calls the deep Q network model for inference, and outputs the frequency, amplitude and timing instructions of ultrasonic vibration to the execution layer; Self-evolution layer: includes an experience replay storage module and a transfer learning module. The experience replay storage module is used to store mold cavity state data, vibration strategies and product quality data during the injection molding production process. The transfer learning module is used to periodically retrain or update deep Q-network models offline or online based on historical production data.

9. The plastic junction box injection filling control system according to claim 8, characterized in that, The number of ultrasonic transducers is 1 to 4, each corresponding to a different vibration region of the mold cavity; The output power of the ultrasonic generator is adjustable from 0 to 1000W, and the vibration frequency range is 20 to 100kHz.

10. The plastic junction box injection filling control system according to claim 8, characterized in that, The sensor for the arrival of the melt front is either a fiber optic detection sensor or a piezoelectric acoustic wave sensor. The sensing layer also includes a cavity displacement sensor, which is used to monitor the cavity deformation in the thick-walled area in real time to calculate the shrinkage depth; the measurement accuracy of the cavity displacement sensor is not less than ±0.5µm.

Images