A real-time quality control and intelligent decision system for additive manufacturing

Abstract

The application discloses a kind of additive manufacturing shaping real-time quality control and intelligent decision system, this system mainly includes distributed monitoring module and centralized decision module.And the construction of this system includes the following steps: S1, construct multimodal sensor system, real-time acquisition powder laying state, molten pool image etc. time series data;S2, design multi-sequence decision neural network MS-DNN, data is handled by three levels of series network layer by layer;S3, establish distributed-centralized decision framework;S4, design combined action mechanism and reward function;S5, multi-agent collaborative training;S6, deploy intelligent decision system, based on training strategy to verify the quality control ability in actual forming process.The application is through millisecond quality monitoring module MS-DNN, distributed-centralized decision architecture, multi-defect suppression reward mechanism trinity, breaks through the bottleneck of poor real-time and insufficient multi-working condition adaptability of traditional system, significantly improves the stability of additive manufacturing process and part forming quality.

Description

Technical Field

[0001] This invention relates to the fields of additive manufacturing and laser processing, and specifically to a real-time quality control and intelligent decision-making system for additive manufacturing forming. Background Technology

[0002] Currently, in multi-laser powder bed additive manufacturing technology, collecting multi-source heterogeneous signals such as molten pool changes, temperature, mechanism motion status, and powder spreading quality during the printing process is crucial for defect detection. While current research has made significant strides in data acquisition systems, it hasn't detailed how to process and analyze the acquired data to maximize its practical value. Furthermore, it lacks timely warnings for abnormal parameters during multi-laser powder bed additive manufacturing and hasn't addressed the system's real-time performance, quality traceability, and intelligent self-diagnosis of anomalies.

[0003] Existing additive manufacturing quality control systems typically only address the measurement of information during various processes. For example, patent CN116237545A discloses a multi-target real-time monitoring device and system for additive manufacturing. This patent achieves cross-scale monitoring through multispectral spectral technology and multi-sensor collaboration: physical isolation of four optical signals is achieved through a three-stage beam splitter; a single galvanometer controls the incident and reflected light collection of the excitation laser, ensuring spatiotemporal consistency of multiple sensors; four-band parallel acquisition: a compressed sensing camera captures the spatter trajectory in the 0.78-0.8μm band to achieve femtosecond-level molten pool morphology monitoring; an infrared thermometer monitors 3000℃ high temperature in the 0.81-1.0μm band; and a photodiode detects the thermal radiation of molten pool spatter particles in the 1.1-3.0μm band (achieving high-speed temperature signal acquisition); an infrared thermal imager simultaneously acquires 0.5-0.54μm visible light and 0.8μm near-infrared light to generate a two-dimensional temperature field distribution. However, this patent does not perform real-time processing and analysis on the monitored multi-source heterogeneous signals, and therefore cannot evaluate or diagnose the monitored quality. Therefore, it is impossible to achieve full-cycle anomaly monitoring and strategy evaluation in the additive manufacturing process. Summary of the Invention

[0004] To address the above challenges, this invention provides a real-time quality control and intelligent decision-making system for additive manufacturing. Through multimodal perception, multi-sequence decision neural network (MS-DNN) quality recognition, and multi-agent reinforcement learning control, it achieves real-time processing and analysis, autonomous decision-making, and closed-loop control of the printing process, thereby coping with complex working condition changes and defect suppression problems and improving the intelligence and stability of the additive manufacturing process.

[0005] A real-time quality control system and intelligent decision-making system for additive manufacturing include: Multiple intelligent agents, each of which performs an independent distributed task in the additive manufacturing process, including powder spreading, atmosphere circulation, forming cylinder movement and scanning strategy adjustment, and melt pool control. The distributed monitoring module includes: A multimodal sensor system is used to collect real-time data on the tasks of powder spreading, atmosphere circulation, forming cylinder movement, and melt pool control. The trained multi-sequence decision neural network (MS-DNN) is connected to the multimodal sensor system to predict the local forming quality state of each agent based on real-time acquired data. The criteria include: solidification runner hump protrusion rate, runner necking rate, runner discontinuity rate, runner flatness rate, and regular runner score. The centralized decision-making module includes: A centralized value network Critic, whose input is the global forming quality state. and the joint actions of all intelligent agents Used to output a value estimate for joint actions ; Multiple distributed policy network actors, each corresponding to an agent, execute a control policy updated after training. The input to the policy is the local shaping quality state independently observed by the corresponding agent. Used to output the control actions of the intelligent agent. ; Among them, the control actions of the multiple intelligent agents The outputs are sent to the corresponding physical controllers: the scraper controller, which adjusts the scraper height KH and scraper speed KV according to the action commands to perform the powder spreading task; the airflow controller, which adjusts the airflow rate AV according to the action commands to perform the air circulation task; the printing controller, which adjusts the scanning spacing SH and scanning angle SA according to the action commands to perform the forming cylinder movement and scanning strategy adjustment task; and the laser controller, which adjusts the laser power P and scanning speed V according to the action commands to perform the melt pool control task.

[0006] The multi-sequence decision neural network (MS-DNN) has a three-level cascaded network structure, comprising: The first-level powder spreading quality learning layer takes powder spreading images as input and is used to identify powder spreading quality. The second-level molten pool change learning layer takes the molten pool change at the next moment after the output of the first-level layer as input and is used to analyze the melting state of the molten pool. The third-level forming evaluation layer takes as input the powder spreading quality data after the first level is completed and the melt pool melting state of the second level, and outputs the forming quality status of the melt channel. ; Each level of the network contains an autoencoder and a decoder. Both the encoder and decoder include a Long Short-Term Memory (LSTM) layer and an attention mechanism module. The decoder uses a teacher-forced input mechanism to input true data. The attention model calculation formula is as follows: In the formula, x i It is the inner product result at the i-th time step, y i It is the attention weight at the i-th time step.

[0007] Furthermore, ,in In the formula, and This indicates the ideal height and ideal width of the melt channel. This represents the standard deviation of the melt height of agent i at time step t. N represents the measured height of agent i at different positions of the melt channel at time step t, and N represents the number of measurement points. This represents the minimum width of the melt channel observed by agent i at time step t; and This indicates that agent i observes the length of the k-th discontinuous melt channel and the total melt channel length at time step t. This represents the number of discontinuous segments in the melt flow counted at time step t; and This represents the actual average height and width of the melt channel for agent i at time step t; denoted as the adjustment coefficient for different agents i.

[0008] Furthermore, the centralized decision-making module is constructed according to the following steps: Step S1.2: Define the prior control strategy for each agent. Step S1.2.1: When the forming quality state has only a single defect, the agent outputs the control action vector according to the prior control strategy. It independently controls the additive manufacturing forming process as an intelligent agent; Step S1.2.2: When concurrent defects exist under the same forming quality condition, the additive manufacturing real-time quality control and intelligent decision-making system first calculates the severity based on the quality index values ​​of each defect, and then generates control priorities accordingly. and normalized regulation priority Get all action vectors Weighted fusion action vector To avoid system oscillations caused by multiple actions running in parallel, the defective action vector with the highest optimization level is identified. Based on the weight of the defect with the highest optimization level Output fine-grained motion vectors Using weighted fusion action vectors and fine motion vectors Constructing combined action vectors According to random actions and combined action vectors Calculate the final output action As an intelligent agent, it independently controls the additive manufacturing forming process; Action vectors It is represented as (P, V, SH, SA, AV, KH, KV), where P represents laser power, V represents scanning speed, SH represents scanning distance, SA represents scanning angle, AV represents gas flow rate, KH represents scraper height, and KV represents scraper speed; the motion vector components are represented by ternary discrete values ​​(-1, 0, 1) to characterize the parameter adjustment direction, +1 indicates increasing the parameter value, 0 indicates maintaining the parameter value, and -1 indicates decreasing the parameter value; Indicates the importance weight of defects. The threshold value for when the i-th type of defect occurs; This represents the control priority value for each defect. It is represented as the sum of the control priority values ​​of all defects. The action vectors for each action in the prior control strategy for forming quality state; This represents the comprehensive suggested direction for all control actions when multiple defects coexist; m represents the mask vector, and ⊙ represents the element-wise product. For critical actions, let m=1, where this dimension is determined by the fine-grained actions. For other non-critical actions, let m=0, and this dimension will remain a weighted fusion of actions. The value remains unchanged; This indicates that it was obtained through Gaussian sampling. Decays over time; and For hyperparameters, Indicates the minimum exploration rate. This represents the initial exploration rate, and t is the number of training rounds. This represents the decay rate.

[0009] Step S1.3: Construct the global reward function The global reward function includes a reward item for the rule-based melting channel. and penalty terms for each defect state sum: in, Indicated as immediate punishment for defects Indicates the defect penalty weight coefficient and Reflecting the importance of abnormal states in the reward function, The threshold value at which the i-th type of defect occurs; This is represented as the instant reward for the rule-based melting path. This represents the reward weight coefficient of the rule-based melting path, and reflect Its importance in the reward function; Step S1.4: Define the centralized value network Critic and multiple distributed policy networks Actor for the centralized decision-making module; Step S1.5: Define the objective function for optimizing the parameters of the distributed policy network Actor; Step S1.6: Perform collaborative training of the centralized value network Critic and multiple distributed policy networks Actor: The centralized value network Critic calculates the value estimation error based on the temporal difference error sampling experience and updates its parameters to minimize the value estimation error; the distributed policy networks Actor update their parameters based on the gradients provided by the centralized value network Critic to maximize the cumulative expected return; and the target network parameters are synchronized using an update strategy. Step S1.7: Deploy the centralized decision-making module to the physical controller: The scraper controller executes the powder spreading task according to the scraper height KH and scraper speed KV parameters; the air field controller executes the atmosphere circulation according to the atmosphere flow rate AV parameter; the printing controller executes the forming cylinder movement and scanning strategy adjustment according to the scanning gap SH and scanning angle SA parameters; the laser controller executes the melt pool control according to the laser power P and scanning speed V parameters.

[0010] Furthermore, the prior control strategy defines the direction of control action adjustment when different defect states exceed preset thresholds.

[0011] Furthermore, the prior control strategy in step S1.2.1 includes: When the hump of the melting channel is convex Greater than the preset threshold At that time, the action vector is (-1,+1,-1,±1,0,0,0); When the necking rate of the melt channel Greater than the preset threshold At that time, the action vector is (+1,-1,0,0,+1,0,0); When the melt channel discontinuity Greater than the preset threshold At that time, the action vector is (+1,0,-1,±1,0,-1,-1); When the flatness of the melt channel Greater than the preset threshold At that time, the action vector is (+1,-1,0,0,0,-1,0); When the melting channel regularity is scored Greater than the preset threshold At that time, the action vector is (0,0,0,0,0,0,0).

[0012] Furthermore, in step S1.4, the distributed policy network Actor employs two fully connected hidden layers. The hidden layers use the ReLU activation function, and the input is the local shaping quality state obtained by each agent at the current time. The output is a continuous action vector, activated by the Tanh activation function, where the action vector... By strategy Based on the current local forming quality status Generated; The centralized value network Critic consists of two fully connected hidden layers, employing the ReLU activation function, with the current global forming quality state as input. Joint actions of all intelligent agents The output is a value estimate of the control effect of the distributed policy network Actor. .

[0013] Compared with the prior art, the present invention has the following advantages: (1) In the present invention, MS-DNN realizes the millisecond-level output of forming quality indicators, providing timeliness guarantee for closed-loop control; (2) The present invention has adaptive strategy learning capability, realizing automatic response to complex working conditions and defect evolution; (3) The present invention uses multiple agents to jointly optimize strategies under the global evaluation of Critic, realizing coordinated multi-agent collaboration to solve the problem of multiple defects coexisting; (4) The present invention can significantly improve the quality of the final formed parts by suppressing defects and rewarding rule melting channels. Attached Figure Description

[0014] Figure 1 The image shows the developed four-layer deep learning network; Figure 2 The figure shows the implementation of the distributed-centralized decision-making using the multi-agent deep deterministic policy gradient algorithm. Figure 3 The diagram shown is of an automatic diagnosis and intelligent processing system. Detailed Implementation

[0015] To make the technical problems solved by this invention, the technical solutions adopted, and the technical effects achieved clearer, the technical solutions of this invention will be further described below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only for explaining this invention and not for limiting it. Furthermore, it should be noted that, for ease of description, only the parts related to this invention are shown in the accompanying drawings, not all of them.

[0016] A real-time quality control and intelligent decision-making system for additive manufacturing is applied to the aforementioned laser powder bed melting (PBF-LB) process, comprising a distributed monitoring module and a centralized decision-making module.

[0017] The distributed monitoring module uses a sensor integration system to collect real-time data from each independent distributed task in the additive manufacturing process. At the same time, it uses a trained multi-sequence decision neural network to predict the forming quality status based on the real-time data.

[0018] The centralized decision-making module includes a pre-trained centralized value network (Critic) configured for all agents and a pre-trained distributed policy network (Actor) assigned to each agent. The distributed policy network (Actor) is responsible for outputting optimal continuous control commands and reward information based on its local observations of the forming quality state, and shares a perception of the global forming quality state. Each agent independently executes independent distributed tasks to form actions. The parameters of the distributed policy network (Actor) are optimized by the agents based on their evaluation scores for the actions, which are obtained from the centralized value network (Critic).

[0019] The intelligent agents include a scraper controller, an airflow controller, a printing controller, and a laser controller. Each intelligent agent is responsible for executing an independent distributed task in the additive manufacturing process. The independent distributed tasks executed by each intelligent agent are powder spreading, atmosphere circulation, forming cylinder movement and scanning strategy adjustment, and melt pool control.

[0020] The sensor integration system includes a molten pool monitoring system composed of a multimodal sensor module, a synchronous trigger controller, and an optical imaging component. The multimodal sensor module integrates a high-speed camera, a coaxial high-speed camera, and a hyperspectral camera to acquire image information of the molten pool area from different perspectives and spectral ranges. The multimodal sensors are coordinated and controlled by a unified synchronous trigger controller and data acquisition card to achieve microsecond-level time synchronization acquisition. The acquired images are focused onto the printing area by the optical lens of the optical imaging component to achieve spatial information alignment.

[0021] High-speed cameras are used to monitor the surface temperature distribution of the powder layer in real time, record the temperature change trend of the molten pool, and statistically analyze the temperature history to analyze the change pattern of the powder layer; at the same time, they capture phenomena such as delamination and splashing during the forming process, record the changes, and use them for subsequent quality assessment.

[0022] A coaxial high-speed camera monitors the powder spreading, observes the molten pool position and scanning path in real time, and records the unstable behavior during laser-powder interaction, including turbulence, vapor, splashing, and molten pool collisions. It also simultaneously records dynamic behavior, including laser on / off and galvanometer acceleration / deceleration.

[0023] Hyperspectral cameras are used to monitor the width and height of the molten pool, providing accurate spectral data to help assess the forming quality of the molten pool.

[0024] The scraper controller utilizes a distributed monitoring module with a multi-sequence decision neural network to achieve high-precision identification of powder spreading defects and obtain local observations. The scraper controller's distributed strategy network (Actor) adjusts the powder spreading process based on the established early warning model and local observations. The scraper controller's centralized value network (Critic) evaluates the adjustments to the powder spreading process. Based on the evaluation results of the centralized value network (Critic), the scraper controller's distributed strategy network (Actor) optimizes the powder spreading process, adjusts the scraper action in real time, controls the powder spreading thickness, and changes the scraper speed and height to ensure stable and uniform powder spreading quality.

[0025] The wind field controller employs a distributed monitoring module with a multi-sequence decision neural network to monitor the protective gas flow rate and obtain local observations. The distributed strategy network (Actor) of the wind field controller adjusts the protective gas flow rate based on the established protective atmosphere flow field model that meets the requirements for smoke and splash removal and local observations. The centralized value network (Critic) of the wind field controller evaluates the adjustment of the protective gas flow rate. Based on the evaluation results of the centralized value network (Critic), the distributed strategy network (Actor) optimizes the protective gas flow rate, controls the working status of the dust removal nozzles and return air inlets, and, in conjunction with the air pressure balancing mechanism, maintains a stable protective atmosphere within the laser forming chamber. It dynamically adjusts the wind speed and direction to control molten pool splash, smoke diffusion, and thermal zone stability.

[0026] The printing controller and its distributed monitoring module obtain local observations through a coaxial high-speed camera monitoring system of laser beam and molten pool. The multi-sequence decision neural network of the distributed monitoring module uses a filtering and matching algorithm based on the temporal-spatial information of the molten pool monitoring sequence to monitor the molten pool position and scanning path in real time to achieve local quality diagnosis. The distributed strategy network Actor of the printing controller dynamically adjusts the scanning path density and scanning angle changes based on local observations and preset scanning strategies. The centralized value network Critic of the printing controller evaluates the dynamic adjustments. The distributed strategy network Actor of the printing controller optimizes the dynamic control based on the evaluation, including: adjusting the scanning path density, increasing or decreasing the spacing between adjacent paths, and changing the scanning angle (using 0° / 45° / 90° cross scanning between adjacent layers).

[0027] The laser controller addresses the challenges of intense metal plumes, violent spatter collisions, high-speed molten pool movement, and strong thermal / optical overlapping radiation interference within the laser beam-powder interaction region. A distributed monitoring module utilizes a coaxial high-speed camera monitoring system to obtain local observations of the laser beam and molten pool. The distributed monitoring module's multi-sequence decision neural network employs a filtering and matching algorithm based on the temporal and spatial information of the molten pool monitoring sequence to monitor the molten pool position and scanning path in real time, enabling local quality diagnosis. The laser controller's distributed strategy network (Actor) dynamically adjusts the laser power and scanning rate based on local observations and preset laser power and scanning rate strategies. The laser controller's centralized value network (Critic) evaluates and dynamically adjusts these parameters, and the distributed strategy network (Actor) optimizes the laser power and scanning rate based on the evaluation.

[0028] The multi-sequence decision neural network (MS-DNN) consists of three sequentially connected layers. Each layer's autoencoder and decoder comprises an input layer, an LSTM layer, and an attention model. A teacher forcing model is incorporated into the decoder, feeding ground truth values ​​into the decoding layer to accelerate training. The attention mechanism employs an inner product algorithm and a softmax function. The softmax function expression is as follows: In the formula, x i It is the inner product result at the i-th time step, y i It is the attention weight at the i-th time step.

[0029] The Multi-Sequence Decision Neural Network (MS-DNN) is a network structure proposed in this invention, designed to fuse and model multimodal, multi-stage time-series data in the additive manufacturing process for quality prediction. The network consists of three layers, corresponding to three key stages of the manufacturing process: powder spreading, melt pooling, and forming evaluation.

[0030] The first-level MS-DNN is a powder-spreading quality learning layer. Its input is the powder-spreading image, and its output is the powder-spreading quality. The powder-spreading quality includes high-gray-level molten aggregate protrusions, low-gray-level powder-spreading depressions, powder-spreading streaks caused by scraper damage, and falling gray dot-like dust.

[0031] The second-level MS-DNN is a process change learning layer. Its input is the change of the molten pool at the next time step after the output of the first layer, and its output is the melting state of the molten pool.

[0032] The third-level MS-DNN is an LSTM network. The input is the powder spreading quality data after the first-level network is executed, the melt pool melting state output by the second level, and the output is five forming quality indicators: solidification melt channel hump bulge rate, melt channel necking rate, melt channel discontinuity rate, melt channel flattening rate, and regular melt channel score.

[0033] During the training phase, each distributed policy network (Actor) is responsible for outputting optimal continuous control commands and reward information based on its local observations of the forming quality state, and sharing the perception of the global forming quality state. Each agent executes actions independently, and the centralized value network (Critic) evaluates these actions to maximize the cumulative expected reward, guiding the distributed policy network (Actor) to optimize parameters. At the same time, the centralized value network (Critic) calculates the loss function based on reward information, the global forming quality state, and the joint actions of each agent, and minimizes it to optimize parameters.

[0034] During the execution phase, each additive manufacturing agent independently controls the additive manufacturing action based on its local observations and trained policies within a distributed policy network.

[0035] By combining centralized training and distributed execution through distributed Actor Critic reinforcement learning, the intelligence level and adaptive capability of additive manufacturing forming control are effectively improved. Specifically, the following steps are included: Step 1: Establish an additive manufacturing forming control model with a centralized decision-making module. Step 1.1: Define the forming quality state Forming quality status This includes: solidification runner hump rate, runner necking rate, runner discontinuity rate, runner flatness rate, and regular runner score. The calculation formulas are as follows: (1) (2) (3) (4) (5) (6) In equations (1) to (6), , , , and , respectively, represent the solidification channel hump bulge rate, channel necking rate, channel discontinuity rate, channel flattening rate, and regular channel score observed by agent i at time step t. and This indicates the ideal height and ideal width of the melt channel. This represents the standard deviation of the melt channel height of agent i at time step t, where N represents the measured height of agent i at different positions of the melt channel at time step t, and N represents the number of measurement points of the melt channel. This represents the minimum width of the melt channel observed by agent i at time step t; and Let represent the length of the k-th discontinuous melt channel and the total length of the melt channel observed by agent i at time step t, respectively. This represents the number of discontinuous segments in the melt flow counted at time step t; and These represent the actual average height and width of the melt channel for agent i at time step t, respectively. Let be the adjustment coefficients for different agents i. Based on equations (1)-(6), the global forming quality state of multiple agents in the additive manufacturing process is established as follows: ,in , , , , The local forming quality state observed independently for each agent.

[0036] The ideal value of the melt channel (such as) and In the additive manufacturing process, under specific material systems and process conditions, a large number of formed melt channels are experimentally printed. High-resolution microscopy and image processing methods are used to extract the melt channel contour data. After statistical analysis, the average geometric dimension of the sample with the highest quality score is taken as the ideal value benchmark.

[0037] Step 1.2: Define the prior control strategies for each agent. Step 1.2.1: When the forming quality state has only a single defect, the agent outputs the control action vector according to the prior control strategy. It independently controls the additive manufacturing forming process as an intelligent agent; Step 1.2.2: When concurrent defects exist under the same forming quality condition, the additive manufacturing real-time quality control and intelligent decision-making system first calculates the severity based on the quality index values ​​of each defect, and then generates control priorities accordingly. and normalized regulation priority Obtain the weighted fused action vector of all action vectors. To avoid system oscillations caused by multiple parallel actions, the defective action vector with the highest optimization level is identified. Based on the weight of the defect with the highest optimization level Output fine-grained motion vectors Using weighted fusion action vectors and fine motion vectors Constructing combined action vectors Introducing random actions Final output action As an intelligent agent, it independently controls the additive manufacturing forming process; In the additive manufacturing forming control model, the control strategies of each agent need to dynamically adjust process parameters based on real-time status to collaboratively optimize forming quality and efficiency. The above action vectors... It is represented as (P, V, SH, SA, AV, KH, KV), where P represents laser power, V represents scanning speed, SH represents scanning spacing, SA represents scanning angle, AV represents gas flow rate, KH represents scraper height, and KV represents scraper speed; the motion vector components are represented by ternary discrete values ​​(-1, 0, 1) to indicate the parameter adjustment direction, +1 indicates increasing the parameter value, 0 indicates maintaining the parameter value, and -1 indicates decreasing the parameter value.

[0038] The prior control strategies in step 1.2.1 include: When the hump of the melting channel is convex Greater than the preset threshold This indicates uneven energy distribution in the molten pool or excessive scanning speed, leading to localized bulges on the molten surface. In this case, the action vector is (-1,+1,-1,±1,0,0,0). Control strategies include: reducing laser power, increasing scanning speed, decreasing the spacing between adjacent scans, and changing the scanning angle.

[0039] When the necking rate of the melt channel Greater than the preset threshold When the molten pool is not sufficiently expanded laterally or the powder is not fully fused, the action vector is (+1,-1,0,0,+1,0,0). Control strategies include increasing laser power, decreasing scanning speed, and increasing the protective gas flow rate.

[0040] When the melt channel discontinuity Greater than the preset threshold When this occurs, it indicates a defect in the scanning path planning or an abnormal powder supply. The action vector at this time is (+1,0,-1,±1,0,-1,-1). Control strategies include: increasing laser power at discontinuities, reducing the spacing between adjacent scans, changing the scanning angle, and reducing the scraper speed and height.

[0041] When the flatness of the melt channel Greater than the preset threshold This indicates insufficient melt channel height or low energy input. The action vector at this time is (+1,-1,0,0,0,-1,0). Control strategies include: increasing laser power, decreasing scanning speed, and reducing layer thickness.

[0042] When the melting channel regularity is scored Greater than the preset threshold When this condition is met, it indicates that the current parameter combination can maintain stable shaping. At this time, the motion vector is (0,0,0,0,0,0,0). The control strategy then changes to: keeping core parameters such as laser power and scanning speed constant.

[0043] The preset thresholds are derived from statistical analysis of a large number of actual formed samples and set within an empirical range. Laser power and scanning speed are adjusted by the laser controller; scanning strategy is adjusted by the printing controller, which includes adjusting the scanning path density and changing the scanning angle; protective gas flow rate is adjusted by the air field controller; and scraper speed and height are adjusted by the scraper controller.

[0044] Adjusting priority in step 1.2.2 for: (7) Indicates the importance weight of defects. As a current indicator of local defects, Let be the threshold value at which the i-th type of defect occurs. The defect importance weight is defined by the degree of impact of the defect on the finished product, including: Discontinuous weld channels, because they directly lead to incomplete fusion between molten pools, can significantly affect the density and mechanical properties of parts, and may even become a source of fracture, affecting the integrity and service life of the structure. The weighting is 0.4.

[0045] Necking of the weld channel leads to a reduction in local cross-sectional area, becoming a stress concentration point, which reduces the mechanical properties and fatigue life of the part, with a weighting of 0.3.

[0046] Melt channel hump mainly affects surface quality and the uniformity of powder spreading in subsequent layers, but if severe, it may also cause pores or cracks, with a weighting of 0.2%.

[0047] Flattened melt channels typically only manifest as changes in surface morphology and have a limited impact on overall structural performance, accounting for 0.1% of the weight.

[0048] The normalized regulation priority in step 1.2.2 is as follows: (8) In the formula, This represents the control priority value for each defect. It is represented as the sum of the control priority values ​​of all defects.

[0049] The weighted fusion action vector in step 1.2.2 is: (9) The vector This indicates the overall recommended direction for all control actions when multiple defects coexist. This refers to the action vectors of each action in the prior control strategy for the forming quality state.

[0050] The fine motion vector in step 1.2.2 is: (10) The action vector corresponding to the defect with the highest optimization level; The weight of the defect with the highest optimization level; In step 1.2.2, the combined action vectors are used to weighted fuse the actions. Based on the framework, use fine movements Replace the key movements, and keep the rest of the movements. The median remains unchanged; the combined action vector is constructed as follows: (11) Here, m represents the mask vector, and ⊙ represents the element-wise product. For critical actions, let m=1, in which case this dimension is determined by the fine-grained actions. For other non-critical actions, let m=0, and this dimension will remain a weighted fusion of actions. The value remains unchanged.

[0051] Introducing ε-greedy, the final output action is: (12) Among them, random actions Obtained by Gaussian sampling, It decays over time.

[0052] Attenuation method used: (13) in, and For hyperparameters, Indicates the minimum exploration rate. This represents the initial exploration rate, and t is the number of training rounds. This represents the decay rate. This represents the output of all agents. The control priorities of different agents have already been reflected through the masking mechanism in the action fusion stage. Therefore, the ε-greedy stage only adds random exploration to the joint action vector as a whole, without affecting the priority action output of each agent for key defects.

[0053] Step 1.3: Define the global reward function The agent employs a split reward mechanism, defining an immediate penalty independently for each defect, as follows: (14) In equation (14) This is represented as an immediate penalty for the defect. This represents the defect penalty weight coefficient, reflecting the importance of abnormal states in the reward function. This represents the current global forming quality status. The threshold value at which the i-th type of defect occurs; (15) In formula (15) This represents the immediate reward for the rule-based melting process. This is represented by the reward weight coefficient of the rule-based melting path, reflecting... Importance in the reward function.

[0054] (16) In formula (16) This is a global reward function that, by penalizing abnormal states and rewarding the melting point according to rules, drives the agent to optimize the overall forming quality. Penalizing abnormal states refers to... , , , Greater than season =0, and , , , Less than season =1. Reward Rules for Melting Points Greater than season =1, and Less than season =0.

[0055] Step 1.4: Define the intelligent agent Each additive manufacturing agent employs an Actor-Critic architecture, consisting of a distributed policy network (Actor) and a centralized value network (Critic). The distributed policy network (Actor) uses two fully connected hidden layers, each containing 256 nodes. The hidden layers use the ReLU activation function, and the input is the local forming quality state obtained by each agent at that time. Including 5-dimensional states The calculation formulas are shown in (1)-(6), and the output is a 7-dimensional continuous action vector, using the Tanh activation function. Among them, the action... By strategy Based on the current local forming quality status The generated centralized value network Critic consists of two fully connected hidden layers, each with 512 nodes. It uses the ReLU activation function, and the input is the current global shaping quality state. Joint actions of all intelligent agents The output is a value estimate of the control effect of the distributed policy network Actor. .

[0056] Step 1.5: Define the objective function for optimizing the parameters of the distributed policy network Actor. Agent i executes the current local forming quality state The actions output by the distributed policy network Actor To maximize the cumulative expected return of the centralized value network Critic and continuously update the parameters in the distributed policy network Actor. Actor objective function : (17) (18) Represents the policy function μ i The corresponding objective function with respect to parameters The gradient is the derivative used to optimize policy performance; Represents the policy function Regarding parameters The derivative; Indicates the current global forming quality state Combined action with output Expectations; This indicates the sensitivity of a centralized value network (Critic) to actions. Indicates action It is determined by the policy function Based on the current local forming quality status Generated, with parameters as ; The strategy function is calculated using formula (18). The gradient is used to update the parameters θi of the distributed policy network Actor, enabling the agent to learn to choose the optimal action under different states. At the same time, the loss function Loss of the centralized value network Critic is minimized by formula (19), and the parameters ϕ of the centralized value network Critic are updated to improve the accuracy of the value function.

[0057] (19) The target value for the centralized value network Critic is determined by immediate rewards. Evaluation of the Critic target network Gi μ 'Composition' represents the current global reward function and the global formed quality state. The following joint actions were taken. The expected global return afterward; For agent i, in the current global forming quality state The following joint actions were taken. The immediate reward obtained afterward; the Critic reward value of the centralized value network. Rewards are divided into positive and negative rewards. Represents all possible global forming quality states Joint actions Reward r and the next global forming quality state The sum of the logarithms; β is a discount factor used to balance the importance of immediate rewards and future rewards; Gi μ ′(s′,a1′,…,a N ′) is the Critic target network's view of the next global forming quality state. and the next action a1′,…,a of all agents. N The assessment of ′; a′ k =μ′ k (o k ) indicates the next action a′ k It is determined by the target policy μ′ k Based on the current local state o k Generated.

[0058] Step 2: Train the centralized value network Critic and the distributed policy network Actor 2.1 Initialize training parameters Initialize the distributed policy network Actor parameters θ i And the Critic parameter ϕ for centralized value networks.

[0059] Among them, the distributed policy network Actor (parameter θ) i The Actor network (parameter θi′) is used to generate action instructions for each agent in the current state. The Actor target network (parameter θi′) generates the next action based on the next state sampled from the experience pool and replay pool. The centralized value network Critic (parameter ϕ) is used to evaluate the value of the global forming quality state and all agent actions. The Critic target network (parameter ϕ′) is used to estimate the expected value of future rewards.

[0060] Initialize the environment and obtain the initial global forming quality state. For each step t, each agent determines the current global forming quality state. Calling the strategy function Generate action set Input the action into the environment, execute it, and obtain the next global forming quality state. , Instant global reward r.

[0061] 2.2 Generating Empirical Tuples Define the interaction interface between the agent and the environment, and store the state-action-reward transition data units collected during the interaction between the agent and the environment during the training phase into the experience replay pool in chronological order. Each piece of experience data is uniformly represented in the form of a 5-tuple: ( (Source of experience). Each tuple contains the current global forming quality state output by the MS-DNN network. Joint actions output by the distributed policy network Actor = (a1,…,a N The global reward calculated by the reward function Joint action for environmental enforcement The next global forming quality status after feedback This includes, but is not limited to, information on sources of experience, such as sampling timestamps, controller numbers, defect tags, and current process parameter settings. Recording these sources supports subsequent sample weighting and fault backtracking analysis.

[0062] 2.3 Constructing the training dataset Prioritize experience replay, adjusting sampling probabilities based on sample importance to allow the agent to learn more frequently from experiences that significantly influence its decisions. The temporal difference error (TD-error) for each set of experiences is used as a measure of priority; the larger the error, the higher the priority. Sampling probability is proportional to TD, and batches of data are sampled to construct the training dataset. Each sample includes the current global shaping quality state. Joint actions Real-time global reward information r, next global forming quality status And whether it is in terminal state.

[0063] 2.4 Each agent will sample B data samples from its own experience pool for training.

[0064] 2.5 Update the Critic parameters of the centralized value network According to formula (19), the loss function of the centralized value network Critic is calculated, and the parameter ϕ of the centralized value network Critic is updated by using the gradient descent method through the backpropagation algorithm.

[0065] 2.6 Update the parameters of the distributed policy network Actor. According to formula (20), the policy gradient is calculated, and the parameters of the distributed policy network Actor are updated using the gradient ascent method. This causes the policy function to be adjusted in the direction of increasing expected returns.

[0066] (20) The parameters and weights in the distributed policy network Actor are continuously updated using formula (21). : (twenty one) 2.7 Update target network parameters Update the Actor target network parameters θ' using a soft update strategy. i And the target network parameter ϕ' of the Critic: (twenty two) (twenty three) in, This represents the soft update rate, which is typically set to 0.01. It can be adjusted based on the specific task's training stage and convergence level to control the update speed and avoid policy mutations.

[0067] 2.8 Setting Discount Factors and Exploration Rates During the training phase, to achieve a balance between exploration and exploitation in the control strategy, a fixed discount factor γ and an exploration rate ε with a decay mechanism are used to control the update process of the training strategy. The discount factor γ controls the impact of future rewards on the current strategy learning, and is preferably set to 0.99. The exploration rate gradually decays from an initial value of 1 using formula (13).

[0068] 2.9 Testing Phase After training, the system enters the testing phase to verify the quality control capability of the multi-agent policy network during actual implementation. During testing, the learner parameters are kept fixed, exploration noise is disabled, and each agent outputs actions solely based on the trained policy function, achieving decentralized execution and real-time adjustment of control parameters to complete the system's closed-loop control.

[0069] During the testing process, 316L stainless steel was selected as the material, with powder particle size distribution between 15µm and 45µm. A standard thickness substrate was used to construct various typical sample structures, including solid cubic samples and regular test samples with complex paths. The test samples were printed using two control methods: a printing group without a control strategy and a printing group using the multi-agent fusion control strategy proposed in this invention. Both groups of tests were repeated three times under the same process environment to improve test stability.

[0070] Throughout the testing process, a multimodal sensor module was used to comprehensively monitor and evaluate the printing process and the forming results. High-speed cameras and coaxial high-speed cameras were used to acquire images of the molten pool during the molten pool process, monitoring its dynamic evolution, spatter behavior, and the interaction between the laser and powder. The powder layer status was identified using a high-resolution image acquisition system to identify powder-laying defects. After printing, scanning electron microscopy was used to obtain morphological parameters such as the average height, width, and length of discontinuous sections of the molten track, enabling a comprehensive evaluation of the forming quality.

[0071] Simultaneously, the system inputs the acquired image sequences and sensor data into a pre-trained multi-sequence neural network (MS-DNN) model to predict the molten pool state and forming quality state in real time. It also calculates the mean square error and relative error between the predicted and actual measured values ​​to evaluate the accuracy of the neural network prediction. Quality assessment experiments are conducted on two sets of printed samples, comparing the deviations in width, height, and discontinuities between samples printed using the control strategy described in this patent and those without, to evaluate the control accuracy.

Claims

1. A real-time quality control and intelligent decision-making system for additive manufacturing, characterized in that, include: Multiple intelligent agents, each of which performs an independent distributed task in the additive manufacturing process, including powder spreading, atmosphere circulation, forming cylinder movement and scanning strategy adjustment, and melt pool control. The distributed monitoring module includes: A multimodal sensor system is used to collect real-time data on the tasks of powder spreading, atmosphere circulation, forming cylinder movement, and melt pool control. The trained multi-sequence decision neural network (MS-DNN) is connected to the multimodal sensor system to predict the local forming quality state of each agent based on real-time acquired data. ,in , , , and Representing intelligent agents respectively The solidification run hump bulge rate, run necking rate, run discontinuity rate, run flattening rate, and regular run score observed at time step t. The centralized decision-making module includes: A centralized value network Critic, whose input is the global forming quality state. and the joint actions of all intelligent agents Used to output a value estimate for joint actions ,in For policy functions; Multiple distributed policy network actors, each corresponding to an agent, execute a control policy updated after training. The input to the policy is the local shaping quality state independently observed by the corresponding agent. Used to output the control actions of the intelligent agent. ; Among them, the control actions of the multiple intelligent agents The outputs are sent to the corresponding physical controllers: the scraper controller, which adjusts the scraper height KH and scraper speed KV according to the action commands to perform the powder spreading task; the airflow controller, which adjusts the airflow rate AV according to the action commands to perform the air circulation task; the printing controller, which adjusts the scanning spacing SH and scanning angle SA according to the action commands to perform the forming cylinder movement and scanning strategy adjustment task; and the laser controller, which adjusts the laser power P and scanning speed V according to the action commands to perform the melt pool control task.

2. The system according to claim 1, characterized in that, The multi-sequence decision neural network (MS-DNN) has a three-level cascaded network structure, comprising: The first-level powder spreading quality learning layer takes powder spreading images as input and is used to identify powder spreading quality. The second-level molten pool change learning layer takes the molten pool change at the next moment after the output of the first-level layer as input and is used to analyze the melting state of the molten pool. The third-level forming evaluation layer takes as input the powder spreading quality data after the first level is completed and the melt pool melting state of the second level, and outputs the forming quality status of the melt channel. ; Each level of the network contains an autoencoder and a decoder. Both the encoder and decoder include a Long Short-Term Memory (LSTM) layer and an attention mechanism module. The decoder uses a teacher-forced input mechanism to input true data. The attention model calculation formula is as follows: In the formula, It is the inner product result at the i-th time step. It is the first Attention weights for each time step.

3. The system according to claim 2, characterized in that, ,in In the formula, and This indicates the ideal height and ideal width of the melt channel. Represents intelligent agents The standard deviation of the melt flow height at time step t Represents intelligent agents The measured height at different positions of the melt channel at time step t, where N represents the number of measurement points; Represents intelligent agents Minimum width of the melt channel observed at time step t; and Represents intelligent agents At time step t, observe the length of the k-th discontinuous melt channel and the total melt channel length. This represents the number of discontinuous segments in the melt flow counted at time step t; and Represents intelligent agents At time step The actual average height and width of the melt channel; For different intelligent agents The adjustment coefficient.

4. The system according to claim 2, characterized in that, The centralized decision-making module is constructed according to the following steps: Step S1.2: Define the prior control strategy for each agent; Step S1.3: Construct the global reward function The global reward function includes a reward item for the rule-based melting channel. and penalty terms for each defect state sum: in, Indicated as immediate punishment for defects Indicates the defect penalty weight coefficient and Reflecting the importance of abnormal states in the reward function, The threshold value at which the i-th type of defect occurs; This is represented as the instant reward for the rule-based melting path. This represents the reward weight coefficient of the rule-based melting path, and reflect Its importance in the reward function; Indicates the hump rise rate of the melt channel, Indicates the necking ratio of the melt channel, Indicates the discontinuity rate of the melt channel, Indicates the flatness ratio of the melt channel; Step S1.4: Define the centralized value network Critic and multiple distributed policy networks Actors for the centralized decision-making module; Step S1.5: Define the objective function for optimizing the parameters of the distributed policy network Actor; Step S1.6: Perform collaborative training of the centralized value network Critic and multiple distributed policy networks Actor: The centralized value network Critic calculates the value estimation error based on the temporal difference error sampling experience and updates its parameters to minimize the value estimation error; the distributed policy networks Actor update their parameters based on the gradients provided by the centralized value network Critic to maximize the cumulative expected return; and the target network parameters are synchronized using an update strategy. Step S1.7: Deploy the centralized decision-making module to the physical controller: The scraper controller executes the powder spreading task according to the scraper height KH and scraper speed KV parameters; the air field controller executes the atmosphere circulation according to the atmosphere flow rate AV parameter; the printing controller executes the forming cylinder movement and scanning strategy adjustment according to the scanning gap SH and scanning angle SA parameters; the laser controller executes the melt pool control according to the laser power P and scanning speed V parameters.

5. The system according to claim 4, characterized in that, The prior control strategies for each agent are defined as follows: Step S1.2.1: When the forming quality state has only a single defect, the agent outputs the control action vector according to the prior control strategy. It independently controls the additive manufacturing forming process as an intelligent agent; Step S1.2.2: When concurrent defects exist under the same forming quality condition, the additive manufacturing real-time quality control and intelligent decision-making system first calculates the severity based on the quality index values ​​of each defect, and then generates control priorities accordingly. and normalized regulation priority Get all action vectors Weighted fusion action vector To avoid system oscillations caused by multiple actions running in parallel, the defective action vector with the highest optimization level is identified. Based on the weight of the defect with the highest optimization level Output fine-grained motion vectors Using weighted fusion action vectors and fine motion vectors Constructing combined action vectors According to random actions and combined action vectors Calculate the final output action As an intelligent agent, it independently controls the additive manufacturing forming process; Action vectors It is represented as (P, V, SH, SA, AV, KH, KV), where P represents laser power, V represents scanning speed, SH represents scanning spacing, SA represents scanning angle, AV represents atmosphere flow rate, KH represents scraper height, and KV represents scraper speed; the motion vector components are represented by ternary discrete values ​​(-1, 0, 1) to characterize the parameter adjustment direction, +1 indicates increasing the parameter value, 0 indicates maintaining the parameter value, and -1 indicates decreasing the parameter value; Indicates the importance weight of defects. The threshold value for when the i-th type of defect occurs; This represents the control priority value for each defect. It is represented as the sum of the control priority values ​​of all defects. The action vectors for each action in the prior control strategy for forming quality state; This represents the comprehensive suggested direction for all control actions when multiple defects coexist; m represents the mask vector, and ⊙ represents the element-wise product. For critical actions, let m=1, then the fine-grained actions... For other non-critical actions, let m=0, and maintain the weighted fusion action. The value remains unchanged; This indicates that it was obtained through Gaussian sampling. Decays over time; and For hyperparameters, Indicates the minimum exploration rate. This represents the initial exploration rate, and t is the number of training rounds. For decay rate, Indicates the exploration rate.

6. The system according to claim 4, characterized in that, The prior control strategy defines the direction of control action adjustment when different defect states exceed preset thresholds.

7. The system according to claim 5, characterized in that, The prior control strategy in step S1.2.1 includes: When the hump of the melting channel is convex Greater than the preset threshold At that time, the action vector is (-1,+1,-1,±1,0,0,0); When the necking rate of the melt channel Greater than the preset threshold At that time, the action vector is (+1,-1,0,0,+1,0,0); When the melt channel discontinuity Greater than the preset threshold At that time, the action vector is (+1,0,-1,±1,0,-1,-1); When the flatness of the melt channel Greater than the preset threshold At that time, the action vector is (+1,-1,0,0,0,-1,0); When the melting channel regularity is scored Greater than the preset threshold At that time, the action vector is (0,0,0,0,0,0,0); in, , , , , These represent preset thresholds for the melt flow hump rate, melt flow necking rate, melt flow discontinuity rate, melt flow flatness rate, and melt flow regularity score, respectively.

8. The system according to claim 4, characterized in that, In step S1.4, the distributed policy network Actor uses two fully connected hidden layers. The hidden layers use the ReLU activation function, and the input is the local shaping quality state obtained by each agent at the current time. The output is a continuous action vector, activated by the Tanh activation function, where the action vector... By strategy Based on the current local forming quality status Generated; The centralized value network Critic consists of two fully connected hidden layers, employing the ReLU activation function, with the current global forming quality state as input. Joint actions of all intelligent agents The output is a value estimate of the control effect of the distributed policy network Actor. .

Images