Adaptive control method for composite material additive manufacturing based on multi-source fusion and reinforcement learning

Abstract

The application discloses a kind of based on multi-source fusion and reinforcement learning's composite material additive manufacturing adaptive control method, first in printing process Synchronous acquisition space position, infrared temperature, tension and compression force and multi-source sensor data such as visible light image;Through the lightweight perception module constructed to each road data is parallelly processed and feature extraction, to force data adopts second-order difference to carry out state discrimination, to temperature data adopts multistage threshold segmentation and extracts forming area statistical feature, and reward feature is generated in combination with image recognition and position mutation detection;The multi-dimensional state features after fusion are combined with reward function, and the agent is trained using reinforcement learning algorithm to learn the optimal printing parameter decision strategy;The trained agent is used for real-time control, and the best printing speed and layer thickness instruction are dynamically output to the mechanical arm according to the current state. The application realizes accurate description and real-time adaptive control of complex multi-physical field coupling process, effectively improves printing quality and process stability.

Description

Technical Field

[0001] This invention relates to the fields of additive manufacturing, information processing, and reinforcement learning, and specifically to an adaptive control method for continuous fiber composite additive manufacturing based on multi-source fusion and reinforcement learning. Background Technology

[0002] Continuous fiber composites are composite materials formed by introducing continuous carbon fibers into a thermoplastic resin base. The additive manufacturing process for this material typically involves heating and melting the thermoplastic resin, impregnating it with carbon fibers, depositing it layer by layer, and then cooling and solidifying it on a printing platform to gradually achieve the forming process from points, lines, and surfaces to a solid volume. This process has advantages such as high material utilization and great design freedom. However, in the additive manufacturing process of continuous fiber composites, printing quality problems such as interlayer debonding and incomplete forming often occur, as well as unique faults such as continuous fiber breakage. These forming quality and manufacturing faults are influenced by a complex coupling of multiple factors, including the interaction between the temperature field and mechanical behavior of incompletely cooled regions, the angle and curvature of the printing path, printing speed, and printing layer thickness. Furthermore, experimental data and numerical simulation results are often limited by material properties, equipment parameters, and model assumptions, making it difficult to accurately summarize the nonlinear coupling process of this technology and determine universally optimal printing parameters through a single experiment or single sensor data.

[0003] Reinforcement learning, a machine learning method that learns optimal control strategies through interaction with the environment, can adaptively adjust decisions based on observed states and reward signals, making it suitable for dynamic process control with uncertainties. In the field of additive manufacturing, existing research has attempted to apply reinforcement learning to optimize process parameters. For example, patent document CN116629128A proposes a deep reinforcement learning-based arc additive forming control method. This method generates single temperature field data through computer numerical simulation and uses this as input to a value evaluation network, optimizing the layer parameters through a reward mechanism until the maximum reward is obtained. However, this method relies solely on single physical field (temperature field) data, making it difficult to comprehensively reflect the real state of multi-physics coupling in additive manufacturing. Furthermore, it is based on simulated data and does not address the processing and fusion of real-time data from multiple sensors in actual production lines.

[0004] On the other hand, regarding the application of multi-source information fusion in the additive manufacturing of continuous fiber composite materials, existing patent document CN118219564A discloses a defect identification method based on multi-source information fusion. This method collects various types of data, including infrared temperature images, visible light images, sound signals, pressure signals, and point cloud information. It extracts features through pixel-level and feature-level fusion and uses multiple cascaded deep neural networks for defect classification, then adjusts printing process parameters based on the identification results. However, this method employs a cascaded structure of multiple deep neural networks, resulting in a complex perception module structure, high training difficulty, high computational resource consumption, and a risk of overfitting. In real-time control scenarios, low processing efficiency may affect the system response speed. Furthermore, this patent does not specifically explain how to achieve real-time optimization control of printing parameters based on defect identification results, lacking a complete closed-loop control scheme.

[0005] In summary, the existing technology has the following problems:

[0006] There is a lack of real-time adaptive control methods based on reinforcement learning: the application of reinforcement learning in additive manufacturing is mostly limited to simulation environments or pre-process parameter optimization, and an adaptive printing control method that can make dynamic decisions based on real-time multi-source sensor data has not yet been formed and is applicable to actual production lines.

[0007] Relying on single physical field data makes it difficult to comprehensively describe the process: Existing methods such as CN116629128A only use single temperature field data as reinforcement learning input, which cannot accurately reflect the complex state of multi-physical field interaction such as thermo-mechanical coupling in additive manufacturing, resulting in incomplete environmental description and limited control accuracy.

[0008] There are discrepancies between numerical simulation data and actual physical processes: The computer numerical simulation method is affected by assumptions such as model parameters and material constitutive relations, and its data has uncertainties with the actual forming process, resulting in insufficient generalization ability of the optimization results in practical applications.

[0009] Multi-source information fusion has a complex structure and poor real-time performance: As described in CN118219564A, the existing method uses multiple deep neural networks connected in series to achieve multi-source information fusion and recognition. The module structure is large and the training and inference resources are high, which makes it difficult to meet the requirements of real-time control for processing efficiency and response speed.

[0010] The process parameter optimization and control methods are unclear: Although the existing multi-source fusion methods can achieve defect identification, they do not provide specific real-time optimization and execution mechanisms for printing parameters, resulting in an incomplete closed-loop link from perception to control.

[0011] Therefore, there is an urgent need to propose a control method that can integrate information from multiple sensor sources, has a lightweight sensing module, and combines reinforcement learning to achieve real-time adaptive optimization of printing parameters, so as to improve the forming quality and process stability of continuous fiber composite additive manufacturing. Summary of the Invention

[0012] This invention aims to overcome the shortcomings of existing technologies and provide an adaptive control method for continuous fiber composite additive manufacturing based on multi-source fusion and reinforcement learning. The main objective of this invention is:

[0013] 1. Construct a lightweight and efficient real-time processing and feature fusion module for multi-source sensor information to comprehensively and accurately describe the multi-physics coupling state in the additive manufacturing process.

[0014] 2. Based on the reinforcement learning framework, the agent is trained using fused multi-source data, enabling it to learn and output the optimal printing parameters that match the real-time forming state.

[0015] 3. Achieve closed-loop control from multi-source sensing to real-time decision-making, solving problems such as low control accuracy, poor real-time performance, and weak generalization ability caused by reliance on single physical field data, uncertainty of simulation data, or large sensing modules in existing technologies, thereby improving the forming quality and process stability of continuous fiber composite additive manufacturing.

[0016] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0017] The adaptive control method for composite additive manufacturing based on multi-source fusion and reinforcement learning includes the following steps:

[0018] Step 1: While the continuous fiber composite additive manufacturing process is being printed, multi-dimensional sensor data and current printing parameters are being collected simultaneously; the multi-dimensional sensor data includes at least the spatial position of the print head, infrared temperature field data, tensile and compressive strength measurement data, and visible light image data.

[0019] Step 2: Construct a lightweight perception module to perform parallel processing and feature extraction on real-time acquired multi-dimensional sensor data, and perform feature-level fusion and connection to output environmental state features and reward-related features; specifically including:

[0020] The time series of tensile and compressive force measurement data is subjected to second-order difference processing. The mechanical stability state of the printing process is determined based on the absolute value of the difference result, and a mechanical feature vector is formed.

[0021] Multi-level threshold segmentation and region extraction are performed on infrared temperature field data to obtain feature descriptors that reflect the temperature distribution of the printing nozzle and local forming area, forming a temperature field feature vector.

[0022] The trained image detection model is used to analyze visible light images to identify whether there is external human intervention and output binarized intervention features.

[0023] The information on the abrupt position change of the printhead in the additive manufacturing direction (Z-axis) is transformed into a binary stop intervention feature, and then logically ORed with the binary intervention feature of the visible light image to form the final reward-related feature.

[0024] The original spatial location data, the mechanical feature vector, and the temperature field feature vector are concatenated to form environmental state features;

[0025] Step 3: Based on the environmental state features and reward-related features output in Step 2, and combined with the preset printing parameter action space, build and train a reinforcement learning agent, enabling the agent to learn a decision-making strategy from the environmental state to the optimal printing parameters; specifically:

[0026] Establish a reinforcement learning environment and agent: Define the environmental state features output in step 2 as the state space of the environment, and define the printing parameter adjustment action as the action space of the agent. The printing parameters include printing speed percentage and printing layer thickness.

[0027] Constructing a reward function: The reward function calculates the reward value based on the reward-related features (human intervention signal) output in step 2 and the time interval since the last intervention, guiding the agent to learn to extend the fault-free stable printing time;

[0028] Using the historical data processed in steps 1 and 2 as the offline training dataset, and combining it with the constructed reward function, the agent is trained offline; further fine-tuning and optimization are performed using a small amount of actual hands-on interaction data to obtain the trained agent decision policy network.

[0029] Step 4: During the actual printing process, the environmental state features output in real time by the lightweight sensing module are input into the trained agent; the agent makes a real-time decision on the optimal printing parameters and outputs control signals to the robotic arm for execution, thereby achieving adaptive real-time control.

[0030] Furthermore, the present invention also proposes an electronic device and a readable storage medium:

[0031] An electronic device includes a processor and a memory, the memory being used to store one or more programs; characterized in that: when the one or more programs are executed by the processor, the above-described method is implemented.

[0032] A readable storage medium storing a computer program, characterized in that: when the computer program is executed by a processor, the above-described method is implemented.

[0033] Beneficial effects:

[0034] Compared with the prior art, the present invention has the following significant advantages:

[0035] 1. High perception efficiency and low resource consumption: This invention replaces the complex structure of multiple deep neural networks in series in the prior art by designing a lightweight feature extraction module for parallel processing and performing feature fusion in the final stage. This significantly reduces the training difficulty, computational resource consumption and overfitting risk of the perception module, and improves the efficiency of real-time data processing and system response speed.

[0036] 2. Real-time adaptive control based on real data is achieved: This invention directly uses multi-source sensor data from actual production lines as input, avoiding the model uncertainty and poor versatility problems caused by relying on computer numerical simulation data. Through the interaction and learning between the reinforcement learning agent and the real environment, it can dynamically decide the optimal printing parameters based on the real-time and comprehensive process status, effectively solving the problem of the lack of real-time adaptive control methods in the field of continuous fiber composite additive manufacturing.

[0037] 3. Comprehensive and accurate description of complex processes: This invention comprehensively utilizes multi-dimensional information such as infrared temperature (thermal), tension and compression (mechanical), spatial position (kinematic), and visible light images (visual), overcoming the limitations of single physical field data in describing process states as incomplete and inaccurate, and providing intelligent agents with environmental observations that are closer to the real coupled physical fields.

[0038] 4. Possesses excellent generalization and transfer capabilities: The method framework proposed in this invention does not depend on specific materials or absolute process parameters. The agent can quickly adapt to different materials or slightly different production lines by combining offline learning with on-machine fine-tuning, reducing the time cost and material consumption of process exploration, and improving the practicality and economy of the method.

[0039] 5. A complete "perception-decision-execution" closed loop has been formed: From multi-source data acquisition and lightweight fusion processing to reinforcement learning agent decision-making, and then to real-time issuance and execution of printing parameters, this invention forms a complete adaptive control closed loop, which systematically improves the quality stability and anti-interference capability of the continuous fiber composite additive manufacturing process.

[0040] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0041] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0042] Figure 1This describes the offline data acquisition process in the embodiments of this application;

[0043] Figure 2 This refers to the infrared temperature thermal imager information processing and data feature extraction steps in the embodiments of this application;

[0044] Figure 3 This refers to the tensile and compressive data information processing and feature extraction steps in the embodiments of this application;

[0045] Figure 4 This refers to the visible light image data information processing and feature extraction steps in the embodiments of this application.

[0046] Figure 5 This describes the offline data learning and on-machine fine-tuning optimization process in the embodiments of this application.

[0047] Figure 6 This is the data processing flow of the perception and decision module in this embodiment, from inputting multi-dimensional sensor information to outputting printing process parameters. Detailed Implementation

[0048] The adaptive control method for composite additive manufacturing based on multi-source fusion and reinforcement learning proposed in this invention includes the following steps:

[0049] Step 1: While the continuous fiber composite additive manufacturing process is executing the printing process, multi-dimensional sensor data and current printing parameters are collected simultaneously; the multi-dimensional sensor data includes at least the spatial position of the printing nozzle, infrared temperature field data, tensile and compressive strength measurement data, and visible light image data.

[0050] Step 2: Construct a lightweight perception module to perform parallel processing and feature extraction on real-time multi-dimensional sensor data, and perform feature-level fusion and connection to output environmental state features and reward-related features for reinforcement learning.

[0051] Since tensile and compressive stress measurement data, infrared temperature field data, and visible light image data contain environmental noise or too much non-interesting background data, in order to improve the robustness of this method in actual generation and control, these three types of data collected in real time are subjected to information processing and feature extraction operations.

[0052] Tension and compression measurement data:

[0053] The force measurement values ​​acquired by the tension and compression sensors are cached to obtain a time series of tension and compression measurement data. A second-order difference algorithm is used to obtain the second-order difference results of the tension and compression measurement data. The results are then analyzed based on the absolute value of the second-order difference results. Determine the mechanical state of the print, when This indicates that the mechanical state of the printed image is in a stable state. This indicates that the mechanical state of the printed object is unstable. This indicates that the printed mechanical state is in a highly unstable state. The feature vector composed of the force measurement value and the state identifier is used as the mechanical feature vector.

[0054] The instability and extreme instability arise because, in the continuous fiber composite additive manufacturing process, force measurement values ​​tend to remain stable at fixed values ​​when executing a planned ideal printing path. However, when significant printing quality issues such as unevenness or warping occur in a certain area traversed by the printhead, the printhead comes into direct contact with the forming defects, leading to significant fluctuations in the force measurement values. The more severe the printing quality issues, the larger the absolute value of the second-order difference result of the printhead's mechanical measurement signal.

[0055] Infrared temperature field data:

[0056] Multi-level threshold segmentation and region extraction are performed on the infrared temperature field data to obtain feature descriptors reflecting the temperature distribution of the printing nozzle and local forming area, forming a temperature field feature vector; the specific process is as follows:

[0057] First infrared temperature field data containing the print head and the printed area is extracted from the original infrared temperature field data. The first infrared temperature field data is then segmented according to a set first temperature threshold to obtain the high-temperature region of the print head. The height direction of the original infrared temperature field data is parallel to the axis of the print head.

[0058] The lowest pixel in the height direction of the high-temperature part of the nozzle is taken as the feature position. The data of the high-temperature part of the nozzle is segmented in the height direction according to the feature position to obtain the second infrared temperature field data of the lower part of the high-temperature part of the nozzle, including the printing forming area and the printing platform.

[0059] The second infrared temperature field data is then segmented according to the set second temperature threshold to remove the low-temperature printing platform area that is not of interest, thereby obtaining the third infrared temperature field data that reflects the temperature distribution of the printing nozzle and the local forming area, and obtaining the temperature field feature vector of the third infrared temperature field image; the temperature field feature vector includes the number of pixels in the third infrared temperature field data, the maximum temperature value, the minimum temperature value, the average temperature value, the median temperature value, and the temperature value variance of each pixel.

[0060] Visible light image data:

[0061] Visible light images containing the printhead and the printed area are acquired. Visible light images containing fault intervention behaviors are extracted from these images and labeled as a pre-training set. A visible light image feature detection model is trained using this pre-training set. The input to the visible light image feature detection model is the acquired visible light image containing the printhead and the printed area, and the output is the visible light image feature indicating whether the image contains fault intervention behaviors. These visible light image features are binary data and are denoted as the determination result of the external human intervention behavior.

[0062] The trained image detection model is used to analyze visible light images to identify whether there is external human intervention and output binarized intervention features.

[0063] The positional abrupt change information of the printhead in the additive manufacturing direction (Z-axis) is converted into a binary shutdown intervention feature, and then fused with the visible light image intervention feature by logical OR operation to form the final reward-related feature.

[0064] The binarized shutdown intervention feature is mainly determined by judging the spatial position of the print head in two adjacent data acquisitions. If the change in the spatial position of the print head on the Z-axis is much greater than the normal layer-by-layer increase in height, it is judged that an external manual process intervention has occurred, thus forming the binarized shutdown intervention feature.

[0065] Finally, this step involves concatenating the original spatial location data, the mechanical feature vector, and the temperature field feature vector to form environmental state features; and obtaining reward-related features.

[0066] Step 3: Based on the environmental state features and reward-related features output in Step 2, and combined with the preset printing parameter action space, build and train a reinforcement learning agent, enabling the agent to learn a decision-making strategy from the environmental state to the optimal printing parameters; specifically:

[0067] Step 3.1: Build a reinforcement learning environment and agent: Define the environmental state features output in Step 2 as the state space of the environment, and define the printing parameter adjustment action as the action space of the agent. The printing parameters include printing speed percentage and printing layer thickness.

[0068] Step 3.2: Constructing the reward function: The reward function calculates the reward value based on the reward-related features (human intervention signal) output in Step 2 and the time interval since the last intervention, guiding the agent to learn to extend the fault-free stable printing time; specifically:

[0069]

[0070] In the formula, t is the continuous fault-free printing time since the last intervention event, and the single reward is [0,1]. The attenuation coefficient is... As a penalty constant, C=10 and k=0.1 are taken in this embodiment.

[0071] Step 3.3: Using the historical data processed in Steps 1 and 2 as the offline training dataset, and combining it with the constructed reward function, the agent is trained offline using the near-end policy optimization algorithm; the offline trained agent is deployed to the actual printing equipment, allowing the agent to interact with the real environment in a small amount, and the newly generated interaction data is used to fine-tune the model so that its policy can better adapt to the dynamic characteristics of the actual production line, and finally the trained agent decision policy network is obtained.

[0072] Step 4: During the actual printing process, the environmental state features output in real time by the lightweight sensing module are input into the trained agent; the agent makes a real-time decision on the optimal printing parameters and outputs control signals to the robotic arm for execution, thereby achieving adaptive real-time control.

[0073] The embodiments of the present invention are described in detail below. These embodiments are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0074] In this embodiment, the overall process can be divided into two main stages: the offline training and optimization stage (corresponding to...). Figure 1 , Figure 5 ) and online real-time control stage (corresponding to Figure 6 The first stage involves collecting data and training an intelligent decision-making agent; the second stage utilizes the trained agent to achieve adaptive control.

[0075] Step 1: During the printing process of the continuous fiber composite additive manufacturing process, multi-dimensional sensor data and current printing parameters are simultaneously acquired; the multi-dimensional sensor data includes at least the spatial position of the printing nozzle, infrared temperature field data, tensile and compressive strength measurement data, and visible light image data. For example... Figure 1 As shown.

[0076] Step 1.1: Set the printing task and data acquisition parameters. Select a component with complex geometric features as the printing object, such as a small 20cm × 20cm quadcopter drone support. Generate a standard layer-by-layer printing path file according to its 3D model. Set the data acquisition parameters: the adjustment cycle of the printing parameters is 30 seconds; within each cycle, randomly select a set of printing parameters from the preset discrete motion space, including the printing speed adjustment level (e.g., 0% (hold), 20%, 40%, 60%, 80%, 90%, 100%) and the printing layer thickness correction amount (e.g., +0.1mm, 0mm, -0.1mm). The data sampling frequency of all sensors is uniformly set to 8Hz.

[0077] Step 1.2: Synchronously acquire multi-source data. During the process of the robotic arm executing the above printing path file, the following sensors are simultaneously activated to acquire data and record the execution printing parameters (speed, layer thickness) at each moment.

[0078] Spatial position sensor: Records the coordinates (X, Y, Z) of the printhead tip in three-dimensional space in real time.

[0079] Infrared thermal imager: Aim at the printing nozzle and the formed area to collect temperature field image data.

[0080] Tension / compression sensor: Installed between the printhead and the end of the robotic arm, it measures the contact pressure and fiber tension during the printing process.

[0081] Visible light camera: Monitors the entire printing process and records video streams.

[0082] Step 1.3: Personnel Labeling and Data Storage. Experienced operators monitor the printing process. When issues such as interlayer debonding, fiber breakage, or serious quality defects occur, operators intervene immediately (e.g., pause, clean, reset). These interventions are manually recorded or electronically tagged via specific actions (e.g., pressing an event tag). All collected raw sensor data, printing parameter logs, and personnel intervention event tags are synchronously stored in a local database or file system, forming the original offline dataset.

[0083] Step 2: Construct a lightweight perception module to perform parallel processing and feature extraction on real-time acquired multi-dimensional sensor data, and then perform feature-level fusion and connection to output environmental state features and reward-related features for reinforcement learning. For example... Figure 2 , Figure 3 , Figure 4 and Figure 6 As shown.

[0084] like Figure 2 As shown, the infrared temperature field feature extraction process is as follows:

[0085] Read a frame of data from the infrared thermal imager in real time (e.g., resolution 384×288). Based on pre-calibration, extract a fixed region from the complete image (e.g., pixel range: rows 130 to 288, columns 200 to 358). This region should include the printing nozzle and the heat-affected forming area, yielding the first infrared temperature field data (e.g., 70×70 pixels). Set a first temperature threshold. The first infrared temperature field data is binarized and segmented, retaining data with temperatures higher than [the specified value]. The pixels, primarily corresponding to the high-temperature section of the printhead, are used to obtain data for this region. Within this high-temperature region data, the highest-temperature connected region is identified, and the lowest pixel in the height direction is determined. Using this pixel as a boundary, the high-temperature region data is divided in the height direction, retaining the lower half. This lower half includes the deposited but cooling forming layer and the printing platform background, yielding the second infrared temperature field data. A second temperature threshold is then set. The second infrared temperature field data is segmented to remove data with temperatures lower than [the specified value]. Based on the background of the printing platform, the final result is a third infrared temperature field data that reflects only the temperature distribution of the print head and the local forming area. A set of statistical characteristics of this area are calculated, including: number of effective pixels, highest temperature, lowest temperature, average temperature, median temperature, and temperature variance. These six statistics are combined to form a temperature field feature vector.

[0086] like Figure 3 As shown, the process of extracting mechanical feature vectors is as follows:

[0087] Real-time reading of tensile and compressive force sensor measurements Maintain a sliding window of length 3 in memory to store the force measurement values ​​at three consecutive sampling times. Calculate the second-order difference approximation of the measurement at the current time. And find its absolute value. The mechanical state is determined based on this absolute value: if In this embodiment, If the status code is 0, it indicates "stable"; if In this embodiment, If so, the status code is 5, indicating "unstable"; if If so, the status code is 10, indicating "extremely unstable". The original force measurement values... Combined with the calculated state code, a mechanical feature vector is formed.

[0088] like Figure 4 The process of extracting reward-related features based on visible light image and location fusion is as follows:

[0089] First, a lightweight object detection model (such as YOLOv8n) is trained using approximately 300 images from historical data labeled with "human intervention behavior" (e.g., a hand entering the field of view, tool use). During online runtime, real-time visible light images are input into the model. If an intervention behavior is detected and the confidence level exceeds a threshold (e.g., 0.85), binary features are output. ,otherwise .

[0090] The Z-axis coordinate of the print head is acquired in real time. The difference ΔZ between the current Z-coordinate and the Z-coordinate of the previous sampling period is calculated. If |ΔZ| is much greater than the normal layer thickness (e.g., |ΔZ| is greater than 30mm, and the normal layer thickness is 0.2mm), it is determined that an emergency lift or reset caused by a fault has occurred, and a binary feature is output. ,otherwise .

[0091] Perform a logical OR operation on the two features mentioned above: , This indicates that an incident has occurred that warrants punishment. The process is normal.

[0092] Finally, the original spatial coordinates (X, Y, Z), temperature field feature vector (6-dimensional), and mechanical feature vector (2-dimensional) are directly concatenated to form a comprehensive environmental state feature vector. This vector serves as the input for the reinforcement learning agent's observation of the current manufacturing process.

[0093] Step 3: Based on the environmental state features and reward-related features output in Step 2, and combined with the preset printing parameter action space, build and train a reinforcement learning agent, enabling the agent to learn a decision-making strategy from the environmental state to the optimal printing parameters; specifically:

[0094] Step 3.1: Build a reinforcement learning environment and agent: Define the environmental state features output in Step 2 as the state space of the environment, and define the printing parameter adjustment action as the action space of the agent. The printing parameters include printing speed percentage and printing layer thickness.

[0095] Step 3.2: Constructing the reward function: The reward function calculates the reward value based on the reward-related features (human intervention signal) output in Step 2 and the time interval since the last intervention, guiding the agent to learn to extend the fault-free stable printing time; specifically:

[0096]

[0097] In the formula, t is the continuous fault-free printing time since the last intervention event, the single reward is [0,1], and the manual intervention penalty is -10.

[0098] Step 3.3: As Figure 5 As shown, using the historical data processed in steps 1 and 2 as the offline training dataset, and combining it with the constructed reward function, the agent is trained offline using a proximal policy optimization algorithm. The offline-trained agent is then deployed to an actual printing device. Under safe conditions (e.g., setting limits on parameter variation ranges), the agent is allowed to interact with the real environment in a limited amount of time in an "exploration" mode (e.g., a few printing task cycles). The newly generated interaction data is used to fine-tune the model, allowing its policy to better adapt to the dynamic characteristics of the actual production line, resulting in the trained agent decision policy network.

[0099] Step 4: As Figure 6 As shown, during the actual printing process, the environmental state features output in real time by the lightweight sensing module are input to the trained agent; the agent makes a real-time decision on the optimal printing parameters and outputs control signals to the robotic arm for execution, realizing adaptive real-time printing parameter control for continuous fiber composite additive manufacturing based on multi-source fusion and reinforcement learning.

[0100] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention without departing from the principles and spirit of the present invention.

Claims

1. An adaptive control method for composite additive manufacturing based on multi-source fusion and reinforcement learning, characterized in that: Includes the following steps: Step 1: While performing the printing process in the continuous fiber composite additive manufacturing process, simultaneously collect multi-dimensional sensor data and current printing parameters; The multidimensional sensor data includes at least the spatial position of the printhead, infrared temperature field data, tensile and compressive strength measurement data, and visible light image data. Step 2: Construct a lightweight perception module to perform parallel processing and feature extraction on real-time multi-dimensional sensor data, and perform feature-level fusion and connection to output environmental state features and reward-related features; The time series of tensile and compressive force measurement data is subjected to second-order difference processing. The mechanical stability state of the printing process is determined based on the absolute value of the difference result, forming a mechanical feature vector. Multi-level threshold segmentation and region extraction are performed on the infrared temperature field data to obtain feature descriptors reflecting the temperature distribution of the printing nozzle and local forming area, forming a temperature field feature vector. The specific process is as follows: First infrared temperature field data containing the print head and the printed area is extracted from the original infrared temperature field data. The first infrared temperature field data is then segmented according to a set first temperature threshold to obtain the high-temperature region of the print head. The height direction of the original infrared temperature field data is parallel to the axis of the print head. The lowest pixel in the height direction of the high-temperature part of the nozzle is taken as the feature position. The data of the high-temperature part of the nozzle is segmented in the height direction according to the feature position to obtain the second infrared temperature field data of the lower part of the high-temperature part of the nozzle, including the printing forming area and the printing platform. The second infrared temperature field data is segmented according to the set second temperature threshold, and the low temperature area is removed to obtain the third infrared temperature field data that reflects the temperature distribution of the printing nozzle and the local forming area, and the temperature field feature vector of the third infrared temperature field image is obtained. The temperature field feature vector includes the number of pixels in the third infrared temperature field data, the maximum temperature value, the minimum temperature value, the average temperature value, the median temperature value, and the temperature value variance of each pixel. The reward-related features are obtained through the following process: The trained image detection model is used to analyze visible light images to identify whether there is external human intervention and output binarized intervention features. The information on the abrupt positional changes of the printhead in the layer-by-layer additive direction is converted into a binary shutdown intervention feature; The binarized intervention features and binarized shutdown intervention features are fused by logical OR operation to form the final reward-related features; Step 3: Based on the environmental state features and reward-related features output in Step 2, and combined with the preset printing parameter action space, build and train a reinforcement learning agent so that the agent learns the decision-making strategy from the environmental state to the optimal printing parameters. Step 4: During the actual printing process, the environmental state features output in real time by the lightweight sensing module are input to the trained agent. The agent makes real-time decisions and outputs the optimal printing parameter control instructions to the actuator to achieve adaptive real-time control.

2. The adaptive control method for composite additive manufacturing based on multi-source fusion and reinforcement learning according to claim 1, characterized in that: In step 2, the original spatial location data, the mechanical feature vector obtained from the tensile and compressive measurement data, and the temperature field feature vector obtained from the infrared temperature field data are spliced ​​together to form the environmental state features.

3. An electronic device, comprising a processor and a memory, wherein the memory is used to store one or more programs; characterized in that: When the one or more programs are executed by the processor, the method of any one of claims 1 to 2 is implemented.

4. A readable storage medium storing a computer program, characterized in that: When the computer program is executed by the processor, it implements the method described in any one of claims 1 to 2.

Images