A three-in-one intelligent powder filling mechanism and method based on a large model drive

Abstract

The application discloses a three-in-one intelligent powder filling mechanism and method based on a large model driving, the mechanism comprises a bag warehouse and an automatic bag feeding mechanism, a filling and degassing mechanism, a dust removal and exhaust mechanism, a lower weighing and reweighing compensation module, a vacuum pumping and heat sealing mechanism, a finished product conveying mechanism, a sensing system and an electric control / pneumatic system; the application realizes intelligent monitoring and decision optimization of the powder flow state, the bag body inflation degree and the sealing integrity in the powder filling process by constructing a large language model driving architecture, realizes a fundamental change from traditional programmed control to cognitive intelligent decision, and makes significant progress in control precision, system cooperation, overall production efficiency, maintenance efficiency and system reliability, and provides a replicable and scalable technical paradigm for the industrial automation field.

Description

Technical Field

[0001] This invention belongs to the technical field of intelligent powder filling, specifically relating to a three-in-one intelligent powder filling mechanism and method based on large model driving. Background Technology

[0002] While existing powder packaging equipment is relatively mature in terms of mechanical structure and basic automation, it still lacks intelligent sensing and adaptive control capabilities, which is the main technical bottleneck in the industry. Specifically, traditional negative pressure packaging, screw feeding, and dustproof control systems generally rely on fixed PLC logic and manual parameter settings, failing to automatically adjust packaging parameters based on changes in the density, particle size, humidity, and flowability of different powder materials. When fluctuations occur in the production environment (such as changes in powder state, bag deformation, or airflow disturbances), the system lacks the ability to identify and predict real-time conditions, leading to reduced packaging accuracy and frequent powder blockages or spills. Furthermore, due to the lack of multimodal sensing of the packaging process (such as visual monitoring, weight feedback, and air pressure signal fusion), the equipment cannot comprehensively assess the powder flow state, bag bulging degree, and sealing integrity, resulting in significantly increased decision-making lag and control errors during the packaging process.

[0003] In addition, existing technologies also have the following problems:

[0004] 1) Insufficient control precision and large quantitative error: Traditional weighing systems mostly use single-point feedback, which cannot achieve real-time dynamic correction. When the powder flow rate or pipeline pressure fluctuates, the weighing accuracy decreases, resulting in increased packaging weight deviation.

[0005] 2) Complex structure and high maintenance cost: Existing systems and multi-stage feeding structures generally use mechanical screws, pneumatic valves and multi-channel pipelines. The system structure is complex and difficult to debug. When the equipment fails, the alarm information is simple. Maintenance personnel need to check one by one according to the manual, which cannot quickly locate the root cause. The mean time to repair (MTTR) is long and the equipment cost is high.

[0006] 3) Slow system response and lack of real-time optimization capability: During the packaging process, the equipment's execution units (such as negative pressure valves, feeding screws, and electric clamping mechanisms) operate independently, resulting in high data transmission delays. This makes it impossible to achieve dynamic coordination between different subsystems, thus limiting the overall system response speed and control accuracy.

[0007] 4) Data value not being tapped: A large amount of data generated during equipment operation (such as historical data of each weighing, motor current curves, and vacuum curves) has not been effectively recorded and analyzed, and cannot be used for predictive maintenance and process improvement. Summary of the Invention

[0008] The main objective of this invention is to overcome the problems of insufficient adaptive adjustment, fluctuating quantitative accuracy, poor coupling between dust removal and degassing, and unstable sealing consistency of existing equipment, and to provide a three-in-one intelligent powder filling mechanism and method based on a large model.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] One aspect of the present invention provides a three-in-one intelligent powder filling mechanism based on a large model drive, including a bag storage and automatic bag loading mechanism, a filling and degassing mechanism, a dust removal and exhaust mechanism, a lower weighing and re-weighing compensation module, a vacuuming and heat sealing mechanism, a finished product conveying mechanism, a sensing system, and an electro-pneumatic system.

[0011] The bag storage and automatic bag loading mechanism are used to perform the storage, retrieval, opening, clamping, alignment and height adjustment of packaging bags;

[0012] The filling and degassing mechanism is used to fill the powder into the packaging bag and perform the degassing operation;

[0013] The dust removal and exhaust mechanism suppresses dust;

[0014] The weighing and re-weighing compensation module is used to detect the weight of the powder filling and to perform compensation filling when the weighing error is greater than the threshold.

[0015] The vacuuming and heat-sealing mechanism is used to perform vacuuming and heat-sealing operations on the filled packaging bags;

[0016] The finished product conveying mechanism is used to output the finished product to the downstream process;

[0017] The sensing system is used to acquire multimodal data during the filling process, including bag state images, weighing signals, airflow parameters, and equipment operating status.

[0018] The electronic control / pneumatic system is used to complete the logic control and drive of the three-in-one intelligent powder filling mechanism;

[0019] The electronic control / pneumatic system is equipped with a multimodal data preprocessing system based on a large language model, which generates an optimized control strategy based on the multimodal data acquired by the sensing system.

[0020] As a preferred technical solution, the bag storage and automatic bag loading mechanism includes a bag storage mechanism, a three-axis cylinder, a bag clamping assembly, a positioning plate, and a positioning block;

[0021] The bag dispensing end of the bag storage mechanism is positioned opposite to the bag clamping assembly, which is driven by a cylinder to perform clamping and releasing of the packaging bag.

[0022] The three-axis cylinder is connected to the bag clamping assembly and the positioning plate respectively, and is used to drive the bag clamping assembly and the positioning plate to move in different directions.

[0023] The positioning plate contacts the bag body through the positioning block, completing the positioning and posture adjustment of the bag body before filling.

[0024] As a preferred technical solution, the filling and degassing mechanism includes an upper hopper, a vertical screw feeder head, a servo motor, and a filtration and pulse backflushing unit;

[0025] The upper hopper is located above the frame and is connected to the vertical screw feeder head;

[0026] The servo motor is connected to the vertical screw feeder head via a right-angle reducer.

[0027] The filtration and pulse backflushing unit is connected to the vertical screw feeder head.

[0028] As a preferred technical solution, the dust removal and exhaust mechanism includes a liftable dust removal hood and a follow-up mechanism;

[0029] The liftable dust removal hood is located between the bag clamping assembly and the unloading channel, and its position is adjusted by a follow-up mechanism.

[0030] The follow-up mechanism includes a movable plate bag clamping device mounted on the mounting block. The movable plate bag clamping device is connected in a movable manner through a cylinder connecting block, a pin, a spherical bearing, a shaft, and a seated bearing. The follow-up mechanism forms an air circuit support and interface foundation with the air circuit box in the electro-control / pneumatic system, and is connected to the vacuum air circuit box in the vacuuming and heat-sealing mechanism through the air circuit interface. It is used to coordinately perform dust removal and exhaust operations during the filling and degassing processes.

[0031] As a preferred technical solution, the lower weighing and reweighing compensation module includes a weighing sensor, a weighing box fixing plate, a bag clamping weighing connection, a bag clamping fixing plate, a movable plate, a rotating mounting base, a rotating shaft pin, and a cylinder.

[0032] The weighing sensor is fixedly mounted on the weighing box mounting plate;

[0033] The bag clamping weighing connection, bag clamping fixing plate, and movable plate are connected in sequence to form a gravity transmission path, and the weighing and compensation operations are realized through the rotating mounting base, rotating shaft pin and cylinder.

[0034] As a preferred technical solution, the vacuuming and heat-sealing mechanism includes a vacuum chamber, a vacuuming needle device, a vacuum air circuit box, a heat-sealing assembly, a sliding door panel, a lifting mechanism, and a drive structure.

[0035] The vacuum chamber and the sliding door panel form an opening and closing structure; the vacuum air circuit box is connected to the vacuum chamber through a vacuum needle port device for performing vacuuming operations.

[0036] The heat sealing assembly is disposed inside the vacuum chamber and connected to the drive structure for performing the heat sealing operation.

[0037] The lifting mechanism is used to perform lifting operations.

[0038] As a preferred technical solution, the sensing system includes a visual camera, a weighing sensor, and a pressure sensor.

[0039] As a preferred technical solution, in the electro-control / pneumatic system, the electro-control system adopts a PLC controller, and the pneumatic system adopts a valve island and cylinders.

[0040] Another aspect of the present invention provides a three-in-one intelligent powder filling method based on a large model, which, using the above-mentioned three-in-one intelligent powder filling mechanism based on a large model, includes the following steps:

[0041] Construct and train a large language model; input the multimodal data acquired in real time by the sensing system into the trained visual CLIP-based feature extraction network to generate an optimized control strategy;

[0042] The electronic / pneumatic system optimizes the control of the powder filling stage, vacuuming and sealing stage, and dust removal and exhaust stage according to the optimized control strategy.

[0043] As a preferred technical solution, the construction of the large language model specifically includes:

[0044] The input bag state image is encoded into a visual feature vector using a visual CLIP-based feature extraction network, specifically: ;in, This indicates that the CLIP model extracts deep spatial features from images; The linear projection layer represents the layer used to map the deep spatial features output by the CLIP model to a dimension d that matches the text embedding of the language model; vector A visual embedding representation that is aligned with the semantics of the text;

[0045] Time series data of input weighing signals and airflow parameters Where T is the time step length, a temporal convolutional network combined with a self-attention mechanism is used to capture dynamic patterns, specifically: ;in, It is a temporal convolutional network responsible for extracting local and global temporal dependency features; This is a self-attention layer used to evaluate the importance of features at different time points and generate a weighted composite temporal feature, the final output of which is... It is also a semantic embedding representation of sensor data projected into d dimensions;

[0046] The visual embedding representation and semantic embedding representation are merged and concatenated into a structured text prompt template, specifically as follows: ;in, This indicates a splicing operation. It is a function that converts descriptive text into text embeddings; P is a multimodal cue sequence;

[0047] The large language model generates a control policy sequence C based on the multimodal cue sequence P through regression, specifically: ;in, It is a sequence of control command tokens generated by the model. For sequence length, Indicates a given multimodal cue sequence Under the given conditions, generate the entire control sequence. The joint conditional probability;

[0048] The optimization objective function of the large language model is: ;in, It is a comprehensive reward function, which is a weighted sum of indicators such as fill accuracy, energy consumption, and dust emission. These are the weighting coefficients that balance these objectives.

[0049] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0050] (1) Based on the large language model driven architecture adopted in this invention, the system has achieved a fundamental transformation from traditional programmatic control to cognitive intelligent decision-making, resulting in significant technical effects. In terms of control accuracy, by combining multimodal perception with the deep reasoning ability of the large language model, the system can analyze complex working conditions such as powder state and bag deformation in real time, dynamically optimize filling parameters, significantly reduce the filling weight error from ±2% of the traditional method to within ±0.5%, and at the same time reduce the bag breakage rate by more than 60% and control the dust emission to below 0.1%, thus achieving the goal of refined production.

[0051] In terms of system coordination, the large language model coordinates the three subsystems of filling, negative pressure and dust prevention through a multi-objective optimization algorithm. This breaks through the control conflicts caused by the independent operation of each subsystem in the traditional system, and produces a significant synergistic effect, which improves the system response time by 45%, increases the number of packages per unit time by 30%, reduces energy consumption by 25%, and achieves a qualitative leap in overall production efficiency.

[0052] In terms of maintenance efficiency, a comprehensive health management system has been built based on the intelligent diagnostic system and predictive maintenance model of variational autoencoder. This system enables equipment failure early warning to reach 500 operating hours in advance, reduces average repair time by 60%, and reduces unexpected downtime by 80%, thereby significantly improving the reliability and availability of the equipment.

[0053] In terms of system reliability, the innovative architecture, which uses a large language model as the unified decision-making core, replaces the traditional complex control logic, reducing the amount of control code by 70%, shortening system debugging time by 50%, and increasing the mean time between failures (MTBF) by 3 times. This simplifies the system structure and significantly enhances operational stability. These technological effects together constitute a complete intelligent solution for powder filling equipment, providing a replicable and scalable technological paradigm for the field of industrial automation. Attached Figure Description

[0054] Figure 1 This is a schematic diagram of the structure of a three-in-one intelligent powder filling machine based on a large model driven according to an embodiment of the present invention;

[0055] Figure 2 This is a schematic diagram of the weighing and reweighing compensation module according to an embodiment of the present invention;

[0056] Figure 3 This is a schematic diagram of the box-type vacuuming and heat-sealing mechanism according to an embodiment of the present invention;

[0057] Figure 4 This is a partial structural diagram of the double-channel double-sided heat sealing unit according to an embodiment of the present invention; Figure 5 This is a flowchart illustrating the overall online monitoring feedback and closed-loop optimization process in an embodiment of the present invention.

[0058] In the diagram: 1. Frame A mechanism, 2. Frame B mechanism, 3. Vacuum needle device, 4. Bag storage mechanism, 5. Finished product end conveyor belt, 6. Air circuit box, 7. Cylinder mounting plate A, 8. Connecting piece, 9. Clamping plate connecting plate A, 10. Clamping plate, 11. Clamping plate connecting plate B, 12. L-shaped frame, 13. Cylinder mounting plate B, 14. Three-axis cylinder A, 15. Three-axis cylinder B, 16. Cylinder connecting mounting plate, 17. Shaping mechanism shaft, 18. With vertical bearing seat, 19. Cylinder, 20. Bag clamping assembly, 21. Fisheye connector, 22. Positioning plate, 23. Positioning block A, 24. Positioning block B, 25. Cylinder connecting block, 26. Sliding door panel A. 27. Servo right-angle star reducer; 28. Position indicator + handle; 29. ​​Handwheel shaft; 30. KB20 mounting plate; 31. Lifting shaft; 32. Thin cylinder; 33. 60–90 series servo motor; 34. Vacuum lifting connecting shaft; 35. Ball screw jack; 36. Vacuum air circuit box; 37. Mounting block; 38. Movable plate bag clamping device A; 39. Cylinder connecting block; 40. Pin; 41. Movable plate bag clamping device B; 4 2. Spherical plain bearing; 43. Shaft; 44. Bearing with seat; 45. Weighing box fixing plate; 46. Movable plate A; 47. Movable plate B; 48. Bag clamp fixing plate; 49. Rotary mounting base; 50. Rotary shaft pin; 51. Bag clamp weighing connection; 52. Weighing sensor; 53. Cable chain connector; 54. Reinforcing plate; 55. Cylinder; 56. Vision camera A; 57. Vision camera B; 58. Vision camera C; 59. Vision camera D. Detailed Implementation

[0059] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present application without creative effort are within the scope of protection of the present application.

[0060] This invention revolves around a complete production line process: automatic bagging, vertical spiral filling with parallel degassing, repeated weighing and compensation, box-type vacuuming and double-stage heat sealing, and finished product conveying. The key focus is achieving stable high cleanliness and high precision through modular mechanical structures and parameterized PLC process control. The engineering targets are 80–90 bags / hour per unit and a weighing accuracy of 0–20 g, meeting the requirements for continuous operation and maintainability. The production line is arranged within the factory according to the process sequence, with utilities using AC380V / 50Hz and 0.4–0.6 MPa instrument air, supporting 24-hour and year-round continuous operation. All the above overall objectives and operating conditions are derived from the explicit requirements of the system technical agreement and operation manual.

[0061] Furthermore, this invention also provides a three-in-one intelligent powder filling mechanism and method based on a large language model, to address the technical bottlenecks of existing powder packaging equipment in terms of perception capabilities, decision-making intelligence, and collaborative control. The core of this invention lies in constructing a powder filling system with a large language model as its intelligent hub. Through its powerful contextual understanding, reasoning decision-making, and multi-task coordination capabilities, it achieves global optimization and autonomous decision-making in the filling process, significantly improving the system's intelligence level and overall performance.

[0062] Definitions:

[0063] LLM: Large Language Model;

[0064] CV: Computer Vision.

[0065] CLIP: Contrastive Language-Image Pre-training.

[0066] TCN: Temporal Convolutional Network;

[0067] Few-shot Learning;

[0068] Embedding: Feature Embedding (feature embedding / vector);

[0069] PLC: Programmable Logic Controller.

[0070] PID: Proportional-Integral-Derivative (PID) control.

[0071] HMI: Human-Machine Interface;

[0072] SMC: SMC Corporation (SMC Pneumatic Components);

[0073] SUS304: Stainless Steel 304 (304 stainless steel).

[0074] PTFE: Polytetrafluoroethylene (PTFE / Teflon).

[0075] WC: Tungsten Carbide;

[0076] IP55: Ingress Protection 55 (Protection Level 55);

[0077] MTTR: ​​Mean Time To Repair;

[0078] AC380V / 50Hz: Alternating Current 380V / 50Hz (Three-phase alternating current 380 volts / 50 Hz).

[0079] Example 1:

[0080] like Figure 1 As shown, this embodiment provides a three-in-one intelligent powder filling mechanism based on a large model, which, from front to back, includes a bag hopper and automatic bag loading mechanism, a filling and degassing mechanism, a lower weighing and re-weighing compensation module, a vacuuming and heat sealing mechanism, a dust removal and exhaust mechanism, and a finished product conveying mechanism. It also includes a sensing system and an electro-pneumatic system. Each module / mechanism is structurally arranged on frame A mechanism 1 and frame B mechanism 2, and unfolds sequentially from upstream to downstream in the bag conveying direction. Specifically, the bag hopper and automatic bag loading mechanism correspond to... Figure 1 The bag storage mechanism 4 and its supporting bag opening, bag clamping, positioning and lifting actuators; the filling and degassing mechanism is arranged inside frame A and connected to the bag opening alignment area; the lower weighing and reweighing compensation module corresponds to the area where the weighing sensor 52 and bag clamping weighing connection 51 and other weighing components are located; the vacuuming and heat sealing mechanism is mainly located in frame B, and is implemented in conjunction with the vacuuming needle device 3 and its lifting and guiding connection components; the finished product conveying mechanism corresponds to the finished product end conveyor belt 5.

[0081] 1. Bag storage and automatic bag loading mechanism, used to perform bag storage, bag retrieval, bag opening, clamping, alignment and height adjustment.

[0082] The bag storage and automatic bag loading mechanism are fixedly installed on frame A mechanism 1, including bag storage mechanism 4, three-axis cylinders (including three-axis cylinders A and B, labeled 14 and 15 in the figure), bag clamping assembly 20, positioning plate 22 and positioning blocks (including positioning blocks A and B, labeled 23 and 24 in the figure).

[0083] The bag outlet end of the bag storage mechanism 4 is positioned opposite to the bag clamping assembly 20. The bag clamping assembly 20 is connected to the L-shaped frame 12 via clamping plate connecting plates (including connecting plates A and B, labeled 9 and 11 in the figure) and is driven by cylinder 19 to perform clamping and releasing of the packaging bag.

[0084] The three-axis cylinders (including three-axis cylinders A and B, labeled 14 and 15 in the figure) are respectively mounted on the cylinder mounting plates (including cylinder mounting plates A and B, labeled 7 and 13 in the figure), and their output ends are connected to the bag clamping assembly 20 and the positioning plate 22 to drive the bag clamping assembly 20 and the positioning plate 22 to move in different directions.

[0085] The positioning plate 22 contacts the bag body of the packaging bag through positioning blocks (including positioning blocks A and B, labeled 23 and 24 in the figure), thereby completing the positioning and posture adjustment of the bag body before filling.

[0086] In one or more preferred embodiments, the bag storage mechanism is equipped with bag dispensing / opening suction cups and a bag loading robot to automatically dispense, open, and guide PE bags to the filling nozzle position; the bag clamping mechanism is a clamping structure that can rise and fall according to the bag type and filling stage, and the bag opening is bulged before feeding to smooth out wrinkles and form a stable feeding section. The bag loading, clamping, guiding, and lifting actions are realized by SMC series cylinders and guide components, and key load-bearing and contact parts are made of SUS304 and aluminum alloy to ensure cleanliness and rigidity.

[0087] 2. Filling and degassing mechanism, used to fill powder into packaging bags and perform degassing operation.

[0088] The filling and degassing mechanism includes an upper hopper, a vertical screw feeder head, a servo motor 33, and a filtration and pulse backflushing unit;

[0089] The upper hopper is located above the frame and is connected to the vertical screw feeder head;

[0090] The servo motor 33 is connected to the vertical screw feeder head via a servo right-angle star reducer 27.

[0091] The filtration and pulse backflushing unit is connected to the vertical screw feeder head, forming a structural component that integrates filling and degassing.

[0092] In one or more preferred embodiments, the upper hopper is 120 L, with a 0.03 mm PTFE coating on the inner surface and a 0.2 mm WC reinforcement on the spiral to reduce adhesion and wear. The vertical spiral feeder head integrates a 1 μm filter and pulse backflushing unit, allowing powder filling and gas release to occur in parallel, with downstream dust collection branches forming a closed-loop dust control system. The system presets two filling rates (coarse and fine) in the PLC formula, automatically switching when the target weight is approached, balancing efficiency and accuracy.

[0093] 3. The weighing and re-weighing compensation module is used to detect the weight of the powder filling and to compensate for the filling when the weighing error is greater than the threshold.

[0094] like Figure 2As shown, the lower weighing and reweighing compensation module includes a weighing sensor 52, a weighing box fixing plate 45, a bag clamping weighing connection 51, a bag clamping fixing plate, a movable plate, a rotating mounting base, a rotating shaft pin, and a cylinder.

[0095] The weighing box fixing plate 45 is installed on the frame B mechanism 2, and the weighing sensor 52 is fixedly installed on the weighing box fixing plate 45.

[0096] The bag clamping weighing connection 51, the bag clamping fixing plate 48, and the movable plates (including movable plates A and B, labeled 46 and 47 in the figure) are connected in sequence to form a gravity transmission path, and cooperate with the cylinder 55 through the rotating mounting base 49 and the rotating shaft pin 50 to realize the weighing and compensation operation.

[0097] In one or more preferred embodiments, the weighing unit is arranged in a "bottom weighing" structure, and the weighing signal is fed back to the PLC in real time for dynamic correction. To reduce fluctuations, a "re-weighing / automatic unloading" function is provided: when the deviation exceeds the threshold after coarse weighing, it immediately enters the micro-compensation stage until it falls into the target window, achieving a packaging accuracy of 0~20 g. The weighing frame, gripper linkage, rotation limiter, and SMC drive cylinder form a stable unloading path, avoiding disturbances caused by drops and impacts.

[0098] IV. Dust removal and exhaust mechanism, used to suppress dust and improve cleanliness.

[0099] The dust removal and exhaust mechanism includes a liftable dust removal hood and a follow-up mechanism;

[0100] The liftable dust removal hood is located between the bag clamping assembly 20 and the unloading channel, and its position is adjusted by a follow-up mechanism.

[0101] The follow-up mechanism includes a movable plate bag clamping device (including movable plate bag clamping devices A and B, labeled 38 and 41 in the figure) mounted on the mounting block 37. The movable plate bag clamping devices (including movable plate bag clamping devices A and B, labeled 38 and 41 in the figure) are connected in a movable manner through a cylinder connecting block 39, a pin 40, a spherical bearing 42, a shaft 43, and a seated bearing 44. The follow-up mechanism forms an air circuit support and interface foundation with the air circuit box 6 in the electro-pneumatic system, and is connected to the vacuum air circuit box 36 in the vacuuming and heat sealing mechanism through the air circuit interface. It is used to coordinate with the filling and degassing operations to perform dust removal and exhaust operations during the filling and degassing process.

[0102] In one or more preferred embodiments, throughout the entire feeding process, based on the temporal characteristics of dust generation, the dust removal channel is linked to pulse backflushing to clean the filter layer and maintain suction efficiency. In the middle stage of operation (the period when the peak feeding time and bag bulging change are obvious), the system increases the suction volume and optimizes the distribution, while in the later stage of operation, the air volume is reduced to avoid affecting the balance stability.

[0103] V. Vacuuming and heat-sealing mechanism, used to perform vacuuming and heat-sealing operations on the filled packaging bags.

[0104] like Figure 3 , Figure 4 As shown, the vacuuming and heat-sealing mechanism includes a vacuum chamber, a vacuum needle port device, a vacuum air circuit box 36, a heat-sealing assembly, a sliding door panel, a lifting mechanism, and a drive structure.

[0105] The vacuum chamber is set inside frame B mechanism 2 and forms an opening and closing structure with sliding door panel 26;

[0106] The vacuum air circuit box 36 is connected to the vacuum chamber via a vacuum needle port device for performing vacuuming operations.

[0107] The heat sealing assembly is located inside the vacuum chamber and is connected to the synchronous belt, bearing seat and other drive structures to perform double-sided heat sealing of the packaging bag.

[0108] The lifting mechanism is used to perform lifting operations. Specifically, the lifting plate is connected to the thin cylinder 32 through the lifting shaft 31, and moves up and down with the cooperation of the ball screw jack 35 and the vacuum lifting connecting shaft 34.

[0109] In one or more preferred embodiments, after filling and weighing, the bag enters a box-type vacuum chamber for vacuuming and double-sided heat sealing. This unit does not directly contact the material. After the vacuum level reaches the set value, two independent heat sealing presses are performed. The sealing area uses PID temperature control and is equipped with a shaping pressure plate to ensure consistent heat sealing for different bag thicknesses and materials. The sealing temperature, pressure, and holding time are stored in the form of bag type formulas, which can be retrieved on the HMI for type changes, avoiding quality fluctuations caused by repeated machine adjustments.

[0110] 6. Finished product conveying mechanism, used to output finished products to downstream processes.

[0111] The finished product conveying mechanism includes a finished product end conveyor belt 5, which connects with the downstream unpacking, packing and winding section to form a continuous logistics closed loop from powder bagging to boxed finished products.

[0112] 7. A sensing system for acquiring multimodal data during the filling process, including bag state images, weighing signals, pressure and airflow parameters, and equipment operating status.

[0113] In one or more preferred embodiments, the sensing system includes a visual camera (including visual cameras A, B, C, and D, denoted as 56, 57, 58, and 59 in the figure), a weighing sensor 52, and a pressure sensor.

[0114] 8. Electrical / pneumatic system, used to complete the logic control and drive of the three-in-one intelligent powder filling mechanism.

[0115] The electronic / pneumatic system also includes a touchscreen HMI component for human-machine interaction.

[0116] The electronic control / pneumatic system is equipped with a multimodal data preprocessing system based on a large language model, which generates an optimized control strategy based on the multimodal data acquired by the sensing system.

[0117] Specifically, in the multimodal data preprocessing system based on a large language model, for visual data, a feature extraction network based on visual CLIP is used to encode the bag image into a visual feature vector; for sensor time-series data, a temporal convolutional network is used to extract feature representations.

[0118] For visual data, a CLIP pre-trained model is used. Its core process involves processing the bag image... Encode into a high-dimensional visual feature vector This process can be described as follows:

[0119] ;

[0120] in, The CLIP model represents the extraction of deep spatial features from images. It is a linear projection layer responsible for mapping the features output by CLIP to a dimension d that matches the text embeddings of the language model. (Vector) This refers to a visual embedding representation that is aligned with the semantics of the text.

[0121] For time-series data from sensors such as weight and pressure Where T is the time step length, a temporal convolutional network combined with a self-attention mechanism is used to capture its dynamic patterns. Its encoding process can be represented as:

[0122] ;

[0123] in, It is a temporal convolutional network responsible for extracting local and global temporal dependent features. This is a self-attention layer used to evaluate the importance of features at different time points and generate a weighted composite temporal feature, the final output of which is... It is also a semantic embedding of sensor data projected into d dimensions. Features of all modalities. , and the original device status text description. After standardization, the text is concatenated and injected into a structured text prompt template, forming the final input of the large language model. This fusion process can be formally represented as constructing a multimodal prompt. :

[0124] ;

[0125] here, This indicates a splicing operation. It is a function that converts descriptive text into text embeddings. In this way, non-textual numerical and visual information is "translated" into a sequence of "pseudo-text" that can be directly understood and processed by large language models. All these features, after being standardized, are aligned with the text embedding space of the large language model to form a unified multimodal representation. This design enables the large language model to "understand" the global state of powder filling, including complex working conditions such as real-time deformation of the bag, powder flow state, and system pressure changes, just like a human expert.

[0126] At the decision-making and control level, large language models demonstrate unique advantages in complex decision-making. Based on a deep understanding of the system state, the model can generate optimized control strategies, including those related to filling height, flow velocity curves, and negative pressure parameters. In particular, through its powerful reasoning capabilities, large language models can predict system responses under different control parameters, achieving forward-looking decision optimization. Large language models are based on the multimodal cue sequences they receive. Automatically and regressively generate control strategy sequences This process can be probabilistically expressed as:

[0127] ;

[0128] in, It is a sequence of control command tokens generated by the model. The sequence length is given. Indicates that given a multimodal cue Under the given conditions, generate the entire control sequence. The joint conditional probability. The model, through its internal self-attention mechanism, generates each current token... At that time, all previously generated tokens will be taken into account. and full context hints This ensures the logical coherence and contextual relevance of the output strategy. Furthermore, the core advantage of the large language model lies in its ability to perform joint reasoning, generating collaborative strategies for the three subsystems of filling, negative pressure, and dust prevention. Its optimization objective can be formalized as a multi-objective constrained problem:

[0129] ;

[0130] The model learns to implicitly balance this optimization problem during its policy generation process. It is a comprehensive reward function, which is composed of a weighted sum of indicators such as filling accuracy, energy consumption, and dust emission. These are the weighting coefficients that balance these objectives. The constraints, on the other hand, ensure that all control parameters (such as fill height) are balanced. Rotation speed negative pressure All operate within the safe physical range of the equipment. Internally, for the coordinated control of the three-in-one system, the large language model can comprehensively consider the mutual influence of the filling mechanism, negative pressure system and dust prevention device, and generate globally optimal coordinated control commands, thus solving the problems of control conflicts and low efficiency caused by the independent operation of each subsystem in traditional systems.

[0131] In one or more preferred embodiments, the equipment supports batch, weight, and alarm information recording and tracing on the HMI, and has operation entry points such as one-button clearing, abnormal shutdown, and fault reset. The electrical control system uses a Siemens PLC and a Kunlun Tongtai touch screen, the control cabinet is IP55, equipped with Schneider / Siemens electrical components, and three-color indicator lights indicate the status; the pneumatic system uses SMC / Airtac valve islands and cylinders.

[0132] IX. Other auxiliary installation parts, connectors, or supports are set together with the functional modules they are attached to, and do not constitute separate functional modules. Specifically, they include connecting piece 8, clamping plate 10, cylinder connecting mounting plate 16, shaping mechanism rotating shaft 17, bearing with vertical seat 18, cylinder 19, fisheye joint 21, position indicator and handle 28, handwheel shaft 29, KB20 mounting plate 30, cable chain connector 53, and reinforcing plate 54.

[0133] The connecting piece 8 is connected to the clamping plate 10 via the clamping plate connecting plate A; the cylinder connecting mounting plate 16 is connected to the three-axis cylinder; the shaping mechanism rotating shaft 17, the bearing with vertical seat 18, and the cylinder 19 are connected in sequence; the cylinder 19 is connected to the fisheye connector 21; the position display and handle 28, the handwheel shaft 29, and the KB20 mounting plate 30 are located on the top of the box-type vacuuming and heat sealing mechanism; the drag chain connector 53 and the reinforcing plate 54 are located in the double-track double-sided heat sealing unit.

[0134] By setting up the above scheme:

[0135] At the engineering implementation level, the large language model is deployed as software in the electrical control system of this device, working in conjunction with the PLC controller and touch-screen HMI. The PLC handles real-time sequential control and actuator driving (such as cylinders, valve islands, servo motors, weighing sampling, and interlocking safety). The large language model receives structured status information from visual cameras 56-59, weighing sensors 52, and process signals such as pressure on the upper-level side, and outputs control decisions based on preset recipe boundaries and process objectives (e.g., coarse / fine feeding switching thresholds, screw feeding rate, vacuum triggering conditions, sealing parameter linkage, and anomaly handling strategies). These decision results are written into the PLC recipe or control variables as parameter sets, enabling adaptive adjustment and coordinated control of each sub-module of the device. Specifically, the intelligent control of the large language model mainly functions in the following aspects: First, for the filling stage, it dynamically generates switching strategies and rate parameters for coarse / fine filling by combining weighing feedback and bag opening status information, and the PLC drives the corresponding actuators to complete the feeding action; Second, for the vacuuming and sealing stage, it generates coordinated parameters for vacuuming and heat sealing actions by combining working conditions and cycle time requirements, so that the bag performs double heat sealing after reaching the set vacuum state and ensures sealing consistency; Third, for the dust removal and exhaust stage, it generates dust control linkage strategies based on the filling / degassing status, so that filling, degassing and exhaust form a closed loop coordination, thereby improving cleanliness and operational stability.

[0136] At the overall system architecture level, this invention innovatively uses a large language model as the core decision-maker for the entire system. This large language model is pre-trained and fine-tuned on massive amounts of industrial data, including industrial control instruction sets, powder mechanical property databases, equipment operation logs, and multimodal sensor data, giving it a deep understanding of powder filling processes. The model's input consists of structured multimodal data, including bag state images collected by visual sensors, real-time feedback data from weight sensors, airflow parameters monitored by pressure sensors, and equipment operating status information. Through its powerful sequence modeling and semantic understanding capabilities, the large language model jointly analyzes these multimodal inputs, outputting comprehensive decisions including filling parameter adjustments, equipment collaborative control strategies, and anomaly handling schemes.

[0137] At the perception and understanding level, this invention constructs a multimodal data preprocessing system for large language models. For visual data, a feature extraction network based on visual CLIP is used to encode bag images into visual feature vectors; for sensor time-series data, a temporal convolutional network is used to extract feature representations.

[0138] At the decision-making and control level, large language models demonstrate unique advantages in complex decision-making. Based on a deep understanding of the system state, the model can generate optimized control strategies, including parameters such as fill height, flow velocity curves, and negative pressure. In particular, large language models, through their powerful reasoning capabilities, can predict system responses under different control parameters, achieving forward-looking decision optimization.

[0139] At the real-time optimization level, this invention designs a dynamic adjustment mechanism based on a large language model. The system continuously monitors multi-source signals during the filling process and utilizes the few-shot learning capability of the large language model to adjust control parameters in real time. When changes in powder properties or environmental disturbances are detected, the large language model can quickly generate adaptive control strategies based on its rich knowledge base, ensuring that the system maintains optimal performance under various operating conditions. This online optimization mechanism based on a large language model significantly improves the system's adaptability to complex operating conditions.

[0140] In anomaly handling and maintenance, the large language model plays a crucial role in intelligent diagnosis and maintenance guidance. Through continuous analysis of equipment operating data using the large language model, the system can identify potential equipment failures and performance degradation at an early stage. When anomalies occur, the large language model can not only quickly locate the root cause of the fault but also generate detailed maintenance guidance plans, significantly reducing the system's mean time to repair. Simultaneously, by analyzing historical operating data, the large language model can autonomously identify opportunities for process improvement, achieving continuous self-optimization of the system.

[0141] It should be noted that the system provided in the above embodiments is only an example of the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure can be divided into different functional modules to complete all or part of the functions described above. The system can be applied to a three-in-one intelligent powder filling method based on a large model driven by the following embodiments.

[0142] Example 2:

[0143] In this embodiment, other implementation methods are provided, specifically:

[0144] The process flow follows the sequence of "screening → loading → automatic weighing → dust removal / cleaning → vacuuming → packaging → handling and boxing → palletizing". After power-on self-test, the system loads the formula and automatically completes bag picking, bag opening, bulging, clamping and height alignment, followed by coarse / fine filling and parallel degassing. Subsequently, under the closed-loop control of the lower weighing, re-weighing / material reduction compensation is completed. During this process, the lifting dust collector hood follows the dust sequence and pulse backflushing. Finally, vacuuming and double heat sealing are completed in the box-type vacuum chamber, and the finished product is shaped and sent to the finished product conveying section. This action chain and sequence are presented as a flowchart on the HMI, and the parameter table and alarm logic are visualized and editable, which is convenient for process engineers to optimize and maintain.

[0145] 1. Standard lithium manganese oxide operating conditions: When implementing, the formula can be set according to the standard lithium manganese oxide operating conditions: target net weight 25 kg, coarse addition to 24.90±0.05 kg, fine addition to 25.00±0.01 kg, allowable error ±20g; after vacuuming to the set vacuum degree, perform double heat sealing, the sealing temperature / time / pressure is automatically matched with the bag thickness, the single machine capacity is stable at about 85 bags / hour, and the sampling pass rate is ≥99.5%.

[0146] 2. Changing Bag Type and Height: To change the bag type, simply adjust the bag storage guide and the bag clamp lifting position, and switch the bag type formula (clamping force, filling nozzle insertion depth, sealing parameters) on the touch screen. This allows for quick bag changeovers without significant mechanical modifications. The above formula settings, production capacity, and accuracy range are consistent with the agreement terms and can serve as the basis for tender acceptance.

[0147] Example 3

[0148] In this example, a three-in-one intelligent powder filling method based on a large model is provided, which can be applied to a three-in-one intelligent powder filling mechanism based on a large model as described in the above embodiments, including the following steps:

[0149] Construct and train a large language model; input the multimodal data acquired in real time by the sensing system into the trained visual CLIP-based feature extraction network to generate an optimized control strategy;

[0150] The electronic / pneumatic system optimizes the control of the powder filling stage, vacuuming and sealing stage, and dust removal and exhaust stage according to the optimized control strategy.

[0151] The construction of the large language model specifically involves:

[0152] The input bag state image is encoded into a visual feature vector using a visual CLIP-based feature extraction network, specifically: ;in, This indicates that the CLIP model extracts deep spatial features from images; The linear projection layer represents the layer used to map the deep spatial features output by the CLIP model to a dimension d that matches the text embedding of the language model; vector A visual embedding representation that is aligned with the semantics of the text;

[0153] Time series data of input weighing signals and airflow parameters Where T is the time step length, a temporal convolutional network combined with a self-attention mechanism is used to capture dynamic patterns, specifically: ;in, It is a temporal convolutional network responsible for extracting local and global temporal dependency features; This is a self-attention layer used to evaluate the importance of features at different time points and generate a weighted composite temporal feature, the final output of which is... It is also a semantic embedding representation of sensor data projected into d dimensions;

[0154] The visual embedding representation and semantic embedding representation are merged and concatenated into a structured text prompt template, specifically as follows: ;in, This indicates a splicing operation. It is a function that converts descriptive text into text embeddings; P is a multimodal cue sequence;

[0155] The large language model generates a control policy sequence C based on the multimodal cue sequence P through regression, specifically: ;in, It is a sequence of control command tokens generated by the model. For sequence length, Indicates a given multimodal cue sequence Under the given conditions, generate the entire control sequence. The joint conditional probability;

[0156] The optimization objective function of the large language model is: ;in, It is a comprehensive reward function, which is composed of a weighted sum of indicators such as filling accuracy, energy consumption, and dust emission. These are the weighting coefficients that balance these objectives.

[0157] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0158] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A three-in-one intelligent powder filling mechanism based on a large model drive, characterized in that, It includes a bag storage and automatic bag loading mechanism, a filling and degassing mechanism, a dust removal and exhaust mechanism, a lower weighing and re-weighing compensation module, a vacuuming and heat sealing mechanism, a finished product conveying mechanism, a sensing system, and an electro-control / pneumatic system; The bag storage and automatic bag loading mechanism are used to perform the storage, retrieval, opening, clamping, alignment and height adjustment of packaging bags; The filling and degassing mechanism is used to fill the powder into the packaging bag and perform the degassing operation; The dust removal and exhaust mechanism suppresses dust; The weighing and re-weighing compensation module is used to detect the weight of the powder filling and to perform compensation filling when the weighing error is greater than the threshold. The vacuuming and heat-sealing mechanism is used to perform vacuuming and heat-sealing operations on the filled packaging bags; The finished product conveying mechanism is used to output the finished product to the downstream process; The sensing system is used to acquire multimodal data during the filling process, including bag state images, weighing signals, airflow parameters, and equipment operating status. The electronic control / pneumatic system is used to complete the logic control and drive of the three-in-one intelligent powder filling mechanism; The electronic control / pneumatic system is equipped with a multimodal data preprocessing system based on a large language model, which generates an optimized control strategy based on the multimodal data acquired by the sensing system.

2. The three-in-one intelligent powder filling mechanism based on a large model-driven approach according to claim 1, characterized in that, The bag storage and automatic bag loading mechanism includes a bag storage mechanism, a three-axis cylinder, a bag clamping assembly, a positioning plate, and a positioning block. The bag dispensing end of the bag storage mechanism is positioned opposite to the bag clamping assembly, which is driven by a cylinder to perform clamping and releasing of the packaging bag. The three-axis cylinder is connected to the bag clamping assembly and the positioning plate respectively, and is used to drive the bag clamping assembly and the positioning plate to move in different directions. The positioning plate contacts the bag body through the positioning block, completing the positioning and posture adjustment of the bag body before filling.

3. The three-in-one intelligent powder filling mechanism based on a large model-driven approach according to claim 1, characterized in that, The filling and degassing mechanism includes an upper hopper, a vertical screw feeder head, a servo motor, and a filtration and pulse backflushing unit; The upper hopper is located above the frame and is connected to the vertical screw feeder head; The servo motor is connected to the vertical screw feeder head via a right-angle reducer. The filtration and pulse backflushing unit is connected to the vertical screw feeder head.

4. The three-in-one intelligent powder filling mechanism based on a large model-driven approach according to claim 1, characterized in that, The dust removal and exhaust mechanism includes a liftable dust removal hood and a follow-up mechanism; The liftable dust removal hood is located between the bag clamping assembly and the unloading channel, and its position is adjusted by a follow-up mechanism. The follow-up mechanism includes a movable plate bag clamping device mounted on the mounting block. The movable plate bag clamping device is connected in a movable manner through a cylinder connecting block, a pin, a spherical bearing, a shaft, and a seated bearing. The follow-up mechanism forms an air circuit support and interface foundation with the air circuit box in the electro-control / pneumatic system, and is connected to the vacuum air circuit box in the vacuuming and heat-sealing mechanism through the air circuit interface. It is used to coordinately perform dust removal and exhaust operations during the filling and degassing processes.

5. The three-in-one intelligent powder filling mechanism based on a large model-driven approach according to claim 1, characterized in that, The lower weighing and reweighing compensation module includes a weighing sensor, a weighing box fixing plate, a bag clamping weighing connection, a bag clamping fixing plate, a movable plate, a rotating mounting base, a rotating shaft pin, and a cylinder; The weighing sensor is fixedly mounted on the weighing box mounting plate; The bag clamping weighing connection, bag clamping fixing plate, and movable plate are connected in sequence to form a gravity transmission path, and the weighing and compensation operations are realized through the rotating mounting base, rotating shaft pin and cylinder.

6. The three-in-one intelligent powder filling mechanism based on a large model-driven approach according to claim 1, characterized in that, The vacuuming and heat-sealing mechanism includes a vacuum chamber, a vacuum needle port device, a vacuum air circuit box, a heat-sealing assembly, a sliding door panel, a lifting mechanism, and a drive structure. The vacuum chamber and the sliding door panel form an opening and closing structure; the vacuum air circuit box is connected to the vacuum chamber through a vacuum needle port device for performing vacuuming operations. The heat sealing assembly is disposed inside the vacuum chamber and connected to the drive structure for performing the heat sealing operation. The lifting mechanism is used to perform lifting operations.

7. The three-in-one intelligent powder filling mechanism based on a large model-driven approach according to claim 1, characterized in that, The sensing system includes a vision camera, a weighing sensor, and a pressure sensor.

8. The three-in-one intelligent powder filling mechanism based on a large model-driven approach according to claim 1, characterized in that, In the aforementioned electronic / pneumatic system, the electronic control system uses a PLC controller, while the pneumatic system uses valve islands and cylinders.

9. A three-in-one intelligent powder filling method based on a large model, characterized in that, The application of any one of claims 1-8, a three-in-one intelligent powder filling mechanism based on a large model, comprises the following steps: Construct and train a large language model; input the multimodal data acquired in real time by the sensing system into the trained visual CLIP-based feature extraction network to generate an optimized control strategy; The electronic / pneumatic system optimizes the control of the powder filling stage, vacuuming and sealing stage, and dust removal and exhaust stage according to the optimized control strategy.

10. The three-in-one intelligent powder filling method based on a large model driven according to claim 9, characterized in that, The construction of the large language model specifically involves: The input bag state image is encoded into a visual feature vector using a visual CLIP-based feature extraction network, specifically: ;in, This indicates that the CLIP model extracts deep spatial features from images; The linear projection layer represents the layer used to map the deep spatial features output by the CLIP model to a dimension d that matches the text embedding of the language model; vector A visual embedding representation that is aligned with the semantics of the text; Time series data of input weighing signals and airflow parameters Where T is the time step length, a temporal convolutional network combined with a self-attention mechanism is used to capture dynamic patterns, specifically: ;in, It is a temporal convolutional network responsible for extracting local and global temporal dependency features; This is a self-attention layer used to evaluate the importance of features at different time points and generate a weighted composite temporal feature, the final output of which is... It is also a semantic embedding representation of sensor data projected into d dimensions; The visual embedding representation and semantic embedding representation are merged and concatenated into a structured text prompt template, specifically as follows: ;in, This indicates a splicing operation. It is a function that converts descriptive text into text embeddings; P is a multimodal cue sequence; The large language model generates a control policy sequence C based on the multimodal cue sequence P through regression, specifically: ;in, It is a sequence of control command tokens generated by the model. For sequence length, Indicates a given multimodal cue sequence Under the given conditions, generate the entire control sequence. The joint conditional probability; The optimization objective function of the large language model is: ;in, It is a comprehensive reward function, which is a weighted sum of indicators such as fill accuracy, energy consumption, and dust emission. These are the weighting coefficients that balance these objectives.

Images