A 2um Laser Welding System with Dynamic Parameter Adjustment Based on Complex High Molecular Polymers

Abstract

A 2um laser welding system based on dynamic parameter adjustment of complex polymers belongs to the technical field of polymer material processing and laser manufacturing. It solves the problems that in the existing 2um laser welding technology, due to the diverse material properties and significant dynamic changes in the thermal behavior during the welding process, the welding mode with fixed parameters is difficult to adapt in real time, resulting in difficult control of the molten pool depth, poor welding uniformity, and insufficient joint strength. By monitoring the melting state and interface behavior of the material during the welding process in real time and dynamically adjusting the laser output parameters according to the monitoring results. This method breaks through the limitations of the traditional fixed-parameter mode and can flexibly optimize the welding process parameters according to the changes in material properties and processing states, thus achieving precise control of the welding process.

Description

Technical Field

[0001] The present invention belongs to the technical field of polymer material processing and laser manufacturing, and particularly relates to a 2μm laser welding system based on dynamic parameter adjustment for complex polymers. Background Art

[0002] Due to their excellent properties such as light weight, high strength, corrosion resistance, and insulation performance, polymers have been widely used in high-end manufacturing fields such as aerospace, medical devices, and electronic packaging. However, the complex molecular structure and thermophysical properties of these materials pose many challenges during processing, significantly limiting their utilization rate. Specifically, polymers have long molecular chains and often have highly cross-linked or branched structures, which result in low heat conduction efficiency and non-uniform melting behavior. In addition, these materials usually have a high coefficient of thermal expansion and are prone to thermal deformation or warping during welding, thus affecting the welding quality and the performance of the manufactured parts. At the same time, since these materials are prone to decomposition or degradation under thermal action, their welding window is narrow, and precise control of welding process parameters is required. These characteristics make polymer materials have great potential in practical applications, but their utilization rate is low due to the high processing difficulty.

[0003] To solve the above problems, laser welding technology has become one of the important methods for polymer material processing due to its advantages of non-contact, high precision, and strong controllability. In particular, the application of lasers with a wavelength of 2μm in polymer welding shows unique potential. Compared with traditional 1μm band lasers, 2μm lasers can significantly improve the light energy absorption rate of polymer materials because this band is closer to the characteristic absorption spectrum of polymer materials, thus effectively enhancing the utilization efficiency of laser energy. This characteristic makes 2μm lasers show higher thermal efficiency and processing adaptability in the melting and welding of polymers. However, although 2μm lasers have potential advantages in welding applications, the limitations of their traditional fixed parameter mode still exist. For complex polymers, due to the diverse material properties and significant dynamic changes in thermal behavior during the welding process, the fixed parameter welding mode is difficult to adapt in real time, resulting in problems such as difficult control of the molten pool depth, poor welding uniformity, and insufficient joint strength. This technical bottleneck restricts the popularization and application of 2μm laser welding technology in actual production. Summary of the Invention

[0004] The present invention provides a 2μm laser welding system based on dynamic parameter adjustment for complex polymers to solve the problems of the existing 2μm laser welding technology, where due to the diverse material properties and significant dynamic changes in thermal behavior during the welding process, the fixed parameter welding mode is difficult to adapt in real time, resulting in difficult control of the molten pool depth, poor welding uniformity, and insufficient joint strength.

[0005] The system includes a welding operation unit, a sensor unit, and a host computer. The host computer includes a dynamic imaging detection unit and an intelligent feedback control unit. The welding operation unit performs the welding operation of the polymer. During the welding process, the dynamic imaging detection unit detects the surface morphology characteristics of the molten pool in real time through the sensor unit and feeds back the surface morphology characteristics of the molten pool to the intelligent feedback control unit. The intelligent feedback control unit adjusts the welding parameters in the welding operation unit according to the surface morphology characteristics of the molten pool.

[0006] Furthermore, the sensor unit includes a vision sensor. The vision sensor is used to take high-resolution pictures of the welding process through a high-speed camera and capture the dynamic changes in the surface morphology of the molten pool in real time.

[0007] Furthermore, the welding operation unit includes a 2um fiber laser, an optical isolation lens group, and a dynamic focusing lens group. During welding, the 2um fiber laser is started to generate high-energy 2um laser. This laser is subjected to isolation treatment through the optical isolation lens group in the optical path, and the beam is focused and transmitted through the dynamic focusing lens group.

[0008] Furthermore, the optical isolation lens group includes two wedge prisms, namely the first wedge prism and the second wedge prism. The function of the optical isolation lens group is to ensure that there is no backlight after the 2um laser passes through.

[0009] Furthermore, the dynamic focusing lens group includes two focusing lenses, namely the first focusing lens and the second focusing lens. The function of the dynamic focusing lens group is to compensate for the focusing and divergence of the laser beam. The focal lengths of each focusing lens are different. When the laser beam passes through the dynamic focusing lens group, the light spot is adjusted, so that the energy density on the polymer is different and the light spot size is different, and the welding surface is dynamically adjusted.

[0010] Furthermore, the dynamic imaging detection unit is built based on the FPN feature pyramid network. The dynamic imaging detection unit includes three layers, namely the low layer, the middle layer, and the high layer. The outputs of the three layers are coupled to obtain the final output of the dynamic imaging detection unit.

[0011] In the low layer, there are successively a convolutional layer, an activation layer, a pooling layer, a batch normalization layer, a residual block layer, and a transposed convolutional layer. In the middle layer, there are successively a convolutional layer, an activation layer, a pooling layer, and a feature fusion layer. In the high layer, there are successively a convolutional layer, a non-linear activation layer, a max pooling layer, a residual block layer, a fully convolutional layer, and a feature fusion layer.

[0012] A residual connection is made between the pooling layer and the feature fusion layer in the middle layer, and the residual connection is also connected to the residual block layer in the lower layer to achieve continuous jumping; the transposed convolutional layer in the lower layer is connected to the feature fusion layer in the middle layer through an upsampling layer, and the output of the lower layer is input into the feature fusion layer in the middle layer; the feature fusion layer in the middle layer is connected to the feature fusion layer in the upper layer through an upsampling layer, and the upsampling layer is also connected to the residual block layer in the upper layer to achieve continuous jumping.

[0013] Further, before the features are input into the upper layer, the features are first extracted through a large-kernel convolution operation.

[0014] Further, the intelligent feedback control unit adopts a three-layer coupling architecture composed of an MPC control module, a PPO reinforcement learning module, and an LSTM long-short term memory neural network module;

[0015] The MPC control module calculates and predicts the observed data of the current state, and optimizes the welding parameters in a rolling time domain to predict the welding parameters for the next welding operation;

[0016] The PPO reinforcement learning module uses the future state prediction and current state real-time feedback provided by the MPC control module to adjust the control strategy to ensure that the intelligent feedback control unit can output the optimal welding parameters in an uncertain environment;

[0017] The LSTM long-short term memory neural network module provides a prediction estimate of the future state based on the stored historical state to help the MPC control module more accurately predict the future state;

[0018] The output of the intelligent feedback control unit is: ; where, is the optimal welding parameter output at time; is the welding parameter output of the MPC control module at time; is the welding parameter output of the PPO reinforcement learning module at time, where, is the optimal control strategy learned by the PPO reinforcement learning module at time, is the current state at time; is the hidden state output of the LSTM long-short term memory neural network module at time; , and are the weighting coefficients of the three-layer coupling architecture respectively.

[0019] The beneficial effects of the system of the present invention are:

[0020] The introduction of dynamic parameter adjustment technology has opened up a new path for the 2μm laser welding of polymer materials. By real-time monitoring the melting state and interface behavior of materials during the welding process, and dynamically adjusting the laser output parameters (including power, frequency, spot size, etc.) according to the monitoring results. This method breaks through the limitations of the traditional fixed parameter mode, and can flexibly optimize the welding process parameters according to the changes of material properties and processing states, so as to achieve precise control of the welding process. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is a structural diagram of the welding operation unit in the embodiment of the present invention;

[0022] Figure 2 It is a working flow diagram of the 2um laser welding system with dynamic parameter adjustment based on complex polymer materials in the embodiment of the present invention;

[0023] Figure 3 It is a pyramid structural diagram of the dynamic imaging detection unit in the embodiment of the present invention;

[0024] Figure 4 It is a structural diagram of the dynamic imaging detection unit in the embodiment of the present invention;

[0025] Figure 5 It is a structural diagram of the intelligent feedback control unit in the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0026] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0027] Embodiment 1

[0028] This embodiment provides a 2um laser welding system with dynamic parameter adjustment based on complex polymer materials. The system includes a welding operation unit, a sensor unit and a host computer. The host computer includes a dynamic imaging detection unit and an intelligent feedback control unit. The welding operation unit performs the welding operation of polymer materials. During the welding process, the dynamic imaging detection unit detects the surface morphology characteristics of the molten pool in real time through the sensor unit, and feeds back the surface morphology characteristics of the molten pool to the intelligent feedback control unit. The intelligent feedback control unit adjusts the welding parameters in the welding operation unit according to the surface morphology characteristics of the molten pool.

[0029] The sensor unit includes a vision sensor, and the vision sensor is used to take high-resolution pictures of the welding process through a high-speed camera to capture the dynamic changes of the surface morphology of the molten pool in real time.

[0030] The welding operation unit is as shown in Figure 1 Figure 4, and includes a 2um fiber laser 1, an optical isolation lens group 2, and a dynamic focusing lens group 3. During welding, the 2um fiber laser 1 is started to generate high-energy 2um laser, and the laser is subjected to isolation treatment through the optical isolation lens group 2 in the optical path, and the beam is focused and transmitted through the dynamic focusing lens group 3.

[0031] Compared with the traditional 1μm wavelength laser, the 2μm laser has a higher absorption spectral line matching with the polymer material, which greatly improves the laser absorption efficiency of the material, thereby reducing energy waste and improving the welding efficiency. The 2um fiber laser outputs the laser beam through the optical fiber. The beam transmitted in the optical fiber has a lower transmission loss, and the laser output is stable, which is suitable for long-time and high-efficiency welding operations.

[0032] The optical isolation lens group 2 includes two wedge prisms, namely the first wedge prism M1 and the second wedge prism M2. The function of the optical isolation lens group 2 is to ensure that there is no backward light after the 2um laser passes through.

[0033] The dynamic focusing lens group 3 includes two focusing lenses, namely the first focusing lens M3 and the second focusing lens M4. The function of the dynamic focusing lens group 3 is to compensate for the focusing and divergence of the laser beam. The focal lengths of each focusing lens are different. When the laser beam passes through the dynamic focusing lens group 3, the light spot is adjusted, so that the energy density on the polymer is different and the light spot size is different, and the welding surface is dynamically adjusted.

[0034] The laser beam is transmitted to the dynamic focusing lens group 3 after passing through the optical isolation lens group 2. The dynamic focusing lens group 3 focuses the parallel laser beam onto the welding area to form a laser focus with a high energy density. The control of the beam focus point is crucial. By adjusting the focal length and angle of the focusing lens, it is ensured that the laser beam can be focused to the required welding depth and welding surface during the welding process. The precision control of this step can effectively avoid welding defects caused by inaccurate focus during the welding process, such as local overheating or uneven fusion.

[0035] The working flow chart of the system described in the present invention is as shown in Figure 2As shown in the figure, when the system of the present invention starts to work, it first detects whether the laser focusing state and energy output are normal. If they are not normal, the system is restarted or the dynamic focusing lens group 3 is checked. If they are normal, the vision sensor is started to collect data in real time, and it is judged whether the data is normal. If the data is not normal, the vision sensor is checked. If it is normal, the sensor data is fed back to the host computer, which includes a dynamic imaging detection unit and an intelligent feedback control unit; the dynamic imaging detection unit detects the surface morphology characteristics of the molten pool in real time through the sensor unit and feeds back the surface morphology characteristics of the molten pool to the intelligent feedback control unit, and the intelligent feedback control unit dynamically adjusts the welding parameters in the welding operation unit according to the surface morphology characteristics of the molten pool.

[0036] Furthermore, as an extension of the solution of this embodiment, the sensor unit may further include a temperature sensor and an infrared sensor, and the temperature sensor can monitor the change of the molten pool temperature in real time. If it is detected that the molten pool temperature is too high, the control system will automatically reduce the laser power output to prevent thermal damage or crack generation of the material caused by overheating. When the molten pool temperature is too low or the energy in the welding area is insufficient, the system will automatically increase the laser power to ensure the smooth progress of the welding process.

[0037] Embodiment 2

[0038] This embodiment further limits Embodiment 1 and further introduces the dynamic imaging detection unit, as Figure 3 shown, the dynamic imaging detection unit is built based on the FPN feature pyramid network, and the dynamic imaging detection unit includes three layers, namely the low layer, the middle layer and the high layer.

[0039] As Figure 4 shown in the structure diagram of the dynamic imaging detection unit, the dynamic imaging detection unit is built based on the FPN feature pyramid network, and the dynamic imaging detection unit includes three layers, namely the low layer, the middle layer and the high layer. The outputs of the low layer, the middle layer and the high layer are respectively , and And the outputs of the three layers are coupled to obtain the final output of the dynamic imaging detection unit;

[0040] In the low layer, there are successively a convolutional layer, an activation layer, a pooling layer, a batch normalization layer, a residual block layer and a transposed convolutional layer. In the middle layer, there are successively a convolutional layer, an activation layer, a pooling layer and a feature fusion layer. In the high layer, there are successively a convolutional layer, a non-linear activation layer, a max pooling layer, a residual block layer, a fully convolutional layer and a feature fusion layer;

[0041] A residual connection is made between the pooling layer and the feature fusion layer in the middle layer, and the residual connection is also connected to the residual block layer in the lower layer to achieve continuous jumping; the transposed convolutional layer in the lower layer is connected to the feature fusion layer in the middle layer through the upsampling layer, and the output of the lower layer is input into the feature fusion layer in the middle layer; the feature fusion layer in the middle layer is connected to the feature fusion layer in the upper layer through the upsampling layer, and the upsampling layer is also connected to the residual block layer in the upper layer to achieve continuous jumping.

[0042] Before the features are input into the upper layer, the features are first extracted through large-kernel convolution operations.

[0043] In the convolutional layer of the lower layer, feature extraction is first performed through a feature extraction module, and then detail scanning is performed through convolutional operations to extract visual features. In the feature extraction module, multi-layer convolutions are used to extract more abstract local detail patterns.

[0044] Combined with upsampling and continuous skip connections, the feature fusion in the multi-scale object detection task is effectively optimized. The network extracts preliminary features through convolutional layers. The lower-layer feature maps contain detail information such as edges and textures, which are suitable for small object detection. After being processed by the non-linear activation function and the pooling layer, the spatial resolution of the lower-layer feature maps decreases, enhancing the sensitivity to local changes. The middle layer abstracts local pattern features such as object parts and textures through deeper convolutional layers. Residual blocks and skip connections ensure that the lower-layer features can be directly transmitted to the middle layer, avoiding information loss and promoting information flow between the lower layer and the middle layer. The upper-layer features mainly extract global pattern information such as the overall structure of the object. In order to fuse with the lower-layer features, the upsampling technique is used to increase the spatial resolution of the upper-layer feature maps and align them with the lower-layer feature maps to ensure the effective fusion of multi-scale features. Skip connections establish a direct connection between the lower layer and the upper layer to help the network combine global information and local details. Through this structure, important detail information can be retained when detecting multi-scale objects, while at the same time improving the recognition ability for complex scenes and small objects to enhance the information fusion of different scales.

[0045] As Figure 3 shown, The features extracted for the lower layer, that is, the initial layer of the dynamic imaging detection unit network structure, can capture fine defects and surface textures. The middle-layer features correspond to the middle-layer output of the dynamic imaging detection unit network, mainly extracting local patterns, that is, capturing pattern changes and the local geometric forms of objects within a local area. is the high-level feature, i.e., the high-level output of the dynamic imaging detection unit network. It mainly monitors the overall morphology, the overall layout during the welding process, and the global output. To adjust the dimensions at different layers to make them match, the above function is corrected. Convolutional dimensionality reduction (Conv1×1) is introduced to unify the number of channels to match the algorithm, and an adaptive weight and attention mechanism are added and corrected. The output of the fused network is expressed as:

[0046] ; where 、 and are adaptive weights, which are calculated by the attention mechanism (SA):

[0047] ;

[0048] Among them, in the adaptive weight function and respectively represent the current calculation layer number, and respectively represent the feature information extracted from the th layer and the th layer, represents the feature attention value of the th layer, that is, it represents the importance of the features of this layer. By calculating the adaptive weight of each layer, the importance of each layer's features is dynamically adjusted during multi-layer feature fusion.

[0049] By optimizing the multi-layer network of the dynamic imaging detection unit through the correction function, the feature weights of each layer can be adaptively adjusted according to the feature distribution of the defects, avoiding the loss of molten pool feedback information, improving the accuracy, and providing strong data for the dynamic parameter adjustment system.

[0050] Example 3

[0051] This example further limits Example 1 and further introduces the intelligent feedback control unit. As Figure 5 shown, the intelligent feedback control unit adopts a three-layer coupling architecture composed of an MPC control module, a PPO reinforcement learning module, and an LSTM long and short time series memory neural network module;

[0052] The MPC control module calculates and processes the observed data of the current state and makes predictions to perform rolling horizon optimization on the welding parameters to predict the welding parameters for the next welding operation;

[0053] The PPO reinforcement learning module uses the future state prediction and current state real-time feedback provided by the MPC control module to adjust the control strategy to ensure that the intelligent feedback control unit can output the optimal welding parameters in an uncertain environment;

[0054] The LSTM long- and short-term memory neural network module provides a predictive estimate of future states based on stored historical states, helping the MPC control module to more accurately predict future states;

[0055] The output of the intelligent feedback control unit is: ; where is the optimal welding parameter output at time; is the welding parameter output of the MPC control module at time; is the welding parameter output of the PPO reinforcement learning module at time, where is the optimal control strategy learned by the PPO reinforcement learning module at time, is the current state at time; is the hidden state output of the LSTM long- and short-term memory neural network module at time; , and are the weighting coefficients of the three-layer coupling architecture, respectively.

[0056] Compared with the traditional PID fuzzy control without memory ability, this embodiment proposes a three-layer intelligent control architecture enhancement system with predictability, self-adaptability and long-term memory ability to achieve the best control of welding parameters in the field of laser welding. The proposed three-layer architecture coupling module effectively realizes the intelligent decision-making and precise control of the dynamic system.

[0057] First, as a model-based control strategy, MPC utilizes the dynamic model of the system. Through the observation of the current state and the prediction of future states, it conducts rolling horizon optimization to calculate the optimal control input for a period of time in the future. The core advantage of MPC lies in its ability to handle constraint conditions and dynamically adjust the control input in each control cycle to ensure that the system always maintains optimal performance in a changing environment. At the same time, as a reinforcement learning algorithm, PPO mainly relies on interactions with the environment to continuously optimize the control strategy. PPO updates the policy network based on the current state and the reward signal, thereby gradually improving the control decision and maximizing the cumulative return. In this framework, PPO uses the state prediction and real-time feedback provided by MPC to adjust the control strategy, ensuring that the controller can make optimal choices in an uncertain environment. LSTM is mainly responsible for modeling time-series data in this framework. By storing and retrieving information over time, it captures the long-term dependencies of the system state. LSTM stores historical state information through its memory cell (Cell State) and selectively retains or updates important temporal information through gating mechanisms. The core role of LSTM is to provide a prediction estimate of the future state of the system based on historical data, which helps MPC more accurately predict future state changes and thus improve the control accuracy.

[0058] As Figure 5 shown, first, the image data is fed back through the dynamic imaging monitoring unit. The modified nonlinear MPC optimizes the control sequence for the next N steps, increases the reinforcement learning compensation factor and the LSTM memory error. The purpose of this modification is to more precisely control the degree of reinforcement intelligence and avoid calculation errors under small feedback signals. The optimized formula is as follows:

[0059] ;

[0060] where represents the output of the modified MPC control module. Taking the control of the temperature parameter in the welding parameters as an example:

[0061] represents the system temperature at time step , is the set temperature, indicating that the system reaches the target temperature. describes the deviation degree of the system from the set target and optimizes the control input to minimize the deviation. represents the final control input at time step , which is obtained through the control signal output by weighted fusion in the coupling equation. represents the final control input at time step , reflects the smoothness of the control input. represents the constraint error at time step , Its function is to measure the constraint error. By minimizing this item, it ensures that the system meets all constraint conditions during the optimal control process, avoids violations of system limitations, and improves the overall stability and performance of the system. and are weight parameters used to adjust the importance of different terms in the objective function, control the smoothness of the input, find a balance between the deviation of the system state from the target, and the constraint conditions, and output smoothly.

[0062] At the same time, the PPO and bidirectional LSTM structures learn and optimize parameters such as laser power and scanning speed in real time to adapt to complex working conditions, and synchronously store the time series during the welding process to improve the smoothness and continuity of welding parameter adjustment.

Claims

1. A 2um laser welding system based on dynamic parameter adjustment of complex polymers, characterized in that: The system comprises a welding operation unit, a sensor unit and a host computer, wherein the host computer comprises a dynamic imaging detection unit and an intelligent feedback control unit; the welding operation unit performs a high molecular polymer welding operation, and during the welding process, the dynamic imaging detection unit detects the surface morphology characteristics of the molten pool in real time through the sensor unit, and feeds back the surface morphology characteristics of the molten pool to the intelligent feedback control unit, and the intelligent feedback control unit adjusts the welding parameters in the welding operation unit according to the surface morphology characteristics of the molten pool; The sensor unit includes a visual sensor, which is used to take high-resolution photos of the welding process through a high-speed camera and capture the dynamic changes of the surface morphology of the molten pool in real time; The welding operation unit comprises a 2um fiber laser (1), an optical isolation lens group (2) and a dynamic focusing lens group (3). During welding, the 2um fiber laser (1) is started to generate a high-energy 2um laser, the laser is isolated by the optical isolation lens group (2) in the optical path, and the light beam is focused and transmitted by the dynamic focusing lens group (3); The dynamic imaging detection unit is built on the basis of the FPN feature pyramid network. The dynamic imaging detection unit includes three layers, namely, a low layer, a middle layer and a high layer, and the outputs of the three layers are coupled to obtain the final output of the dynamic imaging detection unit; The low layers include convolutional layers, activation layers, pooling layers, batch normalization layers, residual block layers, and transposed convolutional layers in sequence; the middle layers include convolutional layers, activation layers, pooling layers, and feature fusion layers in sequence; the high layers include convolutional layers, nonlinear activation layers, maximum pooling layers, residual block layers, full convolutional layers, and feature fusion layers in sequence; A residual connection is made between the pooling layer and the feature fusion layer of the middle layer, and the residual connection is also connected to the residual block layer in the lower layer to achieve continuous jumping; the transposed convolution layer of the lower layer is connected to the feature fusion layer of the middle layer through the upsampling layer, and the output of the lower layer is input to the feature fusion layer of the middle layer; the feature fusion layer of the middle layer is connected to the feature fusion layer of the high layer through the upsampling layer, and the upsampling layer is also connected to the residual block layer of the high layer to achieve continuous jumping.

2. The 2um laser welding system based on complex polymer dynamic parameter adjustment according to claim 1 is characterized in that: The optical isolation lens group (2) comprises two wedge-shaped prisms, namely a first wedge-shaped prism (M1) and a second wedge-shaped prism (M2). The function of the optical isolation lens group (2) is to ensure that no return light is generated after the 2 um laser passes through.

3. The 2um laser welding system based on complex polymer dynamic parameter adjustment according to claim 2 is characterized in that: The dynamic focusing lens group (3) comprises two focusing lenses, namely a first focusing lens (M3) and a second focusing lens (M4). The function of the dynamic focusing lens group (3) is to compensate for the focusing and divergence of the laser beam. The focal length of each focusing lens is different. When the laser beam passes through the dynamic focusing lens group (3), the light spot is adjusted, so that the energy density on the high molecular polymer is different, the light spot size is different, and the welding surface is dynamically adjusted.

4. The 2um laser welding system based on complex polymer dynamic parameter adjustment according to claim 3 is characterized in that: Before the features are input to the higher layers, they are first extracted through a large kernel convolution operation.

5. The 2um laser welding system based on complex polymer dynamic parameter adjustment according to claim 4 is characterized in that: The intelligent feedback control unit uses an MPC control module, a PPO reinforcement learning module, and an LSTM long short-term memory neural network module to form a three-layer coupling architecture; The MPC control module calculates and processes the observation data of the current state and predicts it, and optimizes the welding parameters in the rolling time domain to predict the welding parameters for the next welding work; The PPO reinforcement learning module uses the future state prediction and current state real-time feedback provided by the MPC control module to adjust the control strategy to ensure that the intelligent feedback control unit can output the optimal welding parameters under uncertain conditions; The LSTM long short-term memory neural network module provides a prediction estimate of the future state based on the stored historical state, helping the MPC control module to predict the future state more accurately; The output of the intelligent feedback control unit is: ;in, for The optimal welding parameter output at the moment; For MPC control module Welding parameter output at each moment; Strengthen learning module for PPO The welding parameter output at the moment, where Learned for the PPO reinforcement learning module The optimal control strategy at each moment, for The current state at the moment; LSTM long short time series memory neural network module Hidden state output at the moment; , and are the weighted coefficients of the three-layer coupling architecture respectively.

Images