108 results about "Learning controller" patented technology

A human-guided vision-force fused impedance iterative learning control method for a robotic arm, comprising: analyzing a robot-environment interaction dynamics equation, solving a visual servo acceleration model, and making use of the equation to establish a human-robotic arm-environment interaction dynamics model in an image feature space; acquiring an image feature position and speed curve of a human-guided robot completing an assembly task, and using dynamic movement primitives for coding and generalization; and designing an impedance iterative learning controller which uses image feature tracking errors as control input, learning impedance characteristics when the human-guided robot performs a contact operation, identifying unknown contact dynamics under the interaction between the robot and the environment, and counteracting identified contact interference in the feature space, so as to implement a flexible assembly operation. The control method solves the problems in existing assembly operations that human-robotic arm-environment coupling nonlinear dynamics, unknown contact dynamics of intensive contact assembly tasks and poor generalization of assembly scenarios require relearning for different scenarios, etc.

The invention relates to the technical field of data processing, in particular to a backtracking analysis model construction method based on an attack chain, which comprises the following steps that: a kernel layer security agent acquires process, file and network behavior characteristics in a hardware isolation environment, and generates an event tuple; the tensor network pipeline performs three-dimensional decoupling mapping on the tuple into a behavior fingerprint vector, an orthogonalization noise feature and an asymmetric adjacent tensor, and compresses the behavior fingerprint vector, the orthogonalization noise feature and the asymmetric adjacent tensor into a space-time topology tensor block; the reinforcement learning controller constructs a directed acyclic graph based on the tensor blocks, calculates connectivity loss and outputs an event risk score; the dynamic routing engine constructs a decision tree model according to the risk mark, the burst frequency and the correlation entropy, and implements three-level shunting and a multiple simulation system to generate an anti-interference index; and when the deviation between the physical trajectory and the digital model exceeds the tolerance, the closed-loop feedback weight coefficient updates the loss function parameter and adjusts the channel resource weight. And the problem of threat discovery delay caused by attack chain breakage under massive events is solved.

The invention relates to the technical field of BMS and the like, and provides a battery management system based on adaptive digital twinning, a physical layer of the battery management system comprises a battery pack, a sensor network and an edge computing node, and is responsible for data acquisition and preprocessing; the digital twin layer comprises a self-adaptive multi-scale model and a real-time data engine, battery behaviors are dynamically simulated by coupling electrochemical, thermal and aging models, a future state trajectory prediction result is output, and the real-time data engine fuses sensor data, historical data and simulation data to drive model updating; the intelligent decision-making layer comprises a reinforcement learning controller and a fault prediction module which are deployed in a local server, the reinforcement learning controller dynamically optimizes a charging and discharging strategy according to a prediction result of the digital twinborn layer and issues and executes the charging and discharging strategy, and the fault prediction module analyzes multi-source time sequence data based on an LSTM network so as to early warn thermal runaway and short circuit risks in advance. According to the invention, long-term accurate mapping and adaptive adjustment between the battery physical entity and the digital model can be realized.

The invention relates to the technical field of data center thermal management, and discloses a data center liquid cooling precise temperature control method and system, and the method comprises the following steps: pre-analyzing a calculation task to generate a power consumption prediction sequence; combining the power consumption prediction sequence with the real-time state information of the liquid cooling system, calculating the heat cost of the task at each candidate calculation node through a digital twin model, and selecting a target execution node according to the heat cost; inputting the power consumption prediction sequence and real-time state information into a reinforcement learning controller, and generating a composite coordination action containing a flow regulation instruction and a dynamic power consumption upper limit instruction; and cooperatively controlling the micro-fluidic execution array and the chip power consumption management unit on the target execution node according to the composite coordination action. Through prospective power consumption prediction and two-way cooperative control from the heating end and the heat dissipation end, intelligent distribution and accurate temperature control of thermal loads are achieved, temperature overshoot is effectively restrained, and the overall operation energy consumption of the data center is reduced while the system reliability is guaranteed.

The invention relates to the technical field of steel plate hot-dip production, and discloses an intelligent scheduling system for a steel plate hot-dip production process. The system comprises a working condition sensing module which is used for continuously collecting steel belt inlet temperature, zinc liquid component concentration fluctuation data, air knife pressure time sequence parameters and roller system operation state signals of a galvanizing pot area; the feature extraction module receives the working condition data flow, analyzes the steel strip surface temperature field distribution features and the zinc liquid viscosity change rule, and generates a process feature spectrum; the parameter optimization module dynamically adjusts membership function parameters and rule base weight coefficients of the fuzzy reinforcement learning controller according to the process characteristic spectrum, and outputs an optimized plating layer control parameter set; the self-adaptive plating layer control module generates a plating layer thickness control instruction based on the real-time zinc liquid temperature parameter and the optimized plating layer control rule set; and the dynamic roller system scheduling module realizes global cooperative control of the roller system speed based on the instruction. The system can realize intelligent scheduling of the steel plate hot-dip production process.

The invention discloses a reinforcement learning-based spring steel wire drawing control method, which comprises the following steps of: acquiring real-time process parameters to form a process state data sequence; inputting the process state data sequence into the state space model, and constructing a virtual working condition sample; based on the virtual working condition sample, pre-training a reinforcement learning controller and outputting an initial control strategy; inputting the initial control strategy into a lower-layer strategy network, and outputting a wire drawing speed adjusting instruction; the current process state and the wire drawing speed adjusting instruction serve as synchronous input, and an implicit context vector is generated; extracting control strategy characteristics in the edge controller, performing compressed encoding and forming control strategy representation; uploading the control strategy representation to a cloud server, and outputting a unified global control strategy model; and carrying out anomaly detection on the current process state, and if a detected state reconstruction error exceeds an anomaly judgment threshold, triggering safety control logic. The spring steel wire drawing control device realizes spring steel wire drawing control.

The invention discloses a robot motion control optimization method, device, equipment and medium, and the method comprises the steps: inputting task description in a natural language form, robot state information, a historical execution track and environment feedback information into a preset large language model; through the large language model, selecting a target motion strategy of the robot according to the task description; when it is detected that the task of the robot fails or the action of the robot is abnormal, a reward function used for controlling a deep reinforcement learning controller of the robot is reconstructed; and when the state of the robot is abnormal, parameters of the deep reinforcement learning controller are adjusted and optimized, the motion stability, the task completion efficiency and the strategy adaptability of the robot in a high-risk, high-temperature and high-complexity scene are remarkably improved, and the self-adaptive adjusting and optimizing capability and the abnormity recovery capability of the robot in a dynamic environment are enhanced.

The invention relates to the technical field of boiler control, in particular to a multivariable coupling thermal process regulation and control system and method for carbon pollution governing, and the method comprises the steps: collecting multi-source data such as acoustic emission, temperature, humidity and spectrum, and constructing a feature sequence through time mark alignment and wavelet packet enhancement; a heat value characterization quantity is predicted by using a Shenchang differential equation model fused with dynamic gating, a partition equivalent thermal network model is driven on this basis, and accurate prediction of a future time domain temperature field is realized by dynamically correcting thermal resistance and thermal capacity; based on the prediction result, a control instruction is solved through multi-objective optimization under the condition that the active temperature constraint is met; and in combination with heat flow density feedback, a layered reinforcement learning controller is adopted for online compensation of a pre-feedback instruction, and stable and efficient regulation and control of the boiler are achieved.

The invention discloses an intelligent acquisition risk identification system based on deep learning, and the system comprises a time sequence sample construction module which is used for building a time sequence sample sequence; the hierarchical attention structural feature coding module is used for carrying out structural feature coding on the sequential sample sequence by utilizing a hierarchical attention network; the time sequence feature extraction module is used for inputting the structural feature vector into an ETSform model to perform time sequence feature extraction; the hierarchical time sequence collaborative attention adaptive fusion module is used for performing bidirectional attention interaction and dynamically generating a hierarchical weight and a time sequence weight through a meta-learning controller; the improved CatBoost risk identification module is used for outputting a risk score and a risk type label; and the system fusion module is used for summarizing and merging the risk scores and the risk type labels. The method and the device are suitable for merchant transaction risk identification in an acquiring business scene.

A high-speed matching-engine architecture is disclosed that sustains deterministic sub-microsecond latency while processing more than 10 million order messages per second per core on commodity multi-core processors. Orders reside in cache-aligned Data Holder Nodes whose occupancy and price-level boundaries are tracked with constant-time bitmask operations, eliminating pointer-chasing penalties. Per-core huge-page pools, SIMD copy kernels, and lock-free, cache-line-aligned queues further minimize TLB misses and coherence overheads. Overflow is handled by Push Back / Push Forward cascades that relocate the least- or most-prioritized orders between adjoining nodes without violating price-time priority. Node capacities vary monotonically with book depth and are re-tuned online by a lightweight machine-learning controller that maximizes cache-hit probability under changing market micro-structure. The design tightens spreads, raises match-rate revenue, and complies with stringent regulatory latency caps using standard x86-64, Arm, or other architectures.

Systems, methods, software, and devices are disclosed herein related to trajectory computation by way of a memory-based learning (MBL) controller. An MBL controller in various embodiments stores a set of trajectories in memory. The trajectories connect various initial states of a dynamical system with a target state. In addition to the memory, the controller further includes a processor that collects a current state of the dynamical system and determines, using memory-based learning (MBL) on training instances derived from the set of trajectories, a control policy that defines a trajectory connecting the current state of the dynamical system with the target state. The processor controls the dynamical system according to the control policy.

The invention provides a time mark event triggering reinforcement learning multi-agent consistency control method and system, and relates to the technical field of multi-agent cooperative control. Establishing a multi-agent system dynamics model comprising a leader and a plurality of followers on the time mark; based on the state information of the leader and the follower, constructing a tracking error and consistency error dynamic equation between the intelligent agents; a reinforcement learning controller based on a neural network is constructed, and an optimal control strategy and a performance index function are approached; an event triggering mechanism is designed, and the controller is triggered to update only when the local error of the intelligent agent exceeds a dynamic threshold value; by constructing a Lyapunov function on a time scale and solving a linear matrix inequality condition, bounded consistency of the system is ensured, and a Zeno phenomenon is eliminated. Unified modeling and distributed safety control of the time scale multi-agent system in the non-ideal environment are achieved, communication resources are saved, the convergence speed is increased, and the robustness and practicability of the system are enhanced.

The invention discloses a robot self-adaptive control method and system based on deep learning, and relates to the field of robot control, and the method comprises the steps: receiving and preprocessing multi-modal data, obtaining the preprocessed multi-modal data, carrying out the multi-dimensional dynamic evaluation and event triggering judgment, and carrying out the multi-dimensional dynamic evaluation and event triggering judgment. Generating a dynamic evaluation result of the control mode sign and the dynamic feature vector; selecting a deep learning controller network structure according to a control mode mark in a dynamic evaluation result, performing calculation based on the preprocessed multi-modal data and the dynamic feature vector, and outputting a preliminary joint torque instruction; and the initial joint torque instruction and the preprocessed multi-modal data are input into a micro safety barrier layer, the initial joint torque instruction is corrected according to preset robot safety constraints, and a safety joint torque instruction is generated. According to the method, the parameters of the controller are updated on line based on data such as actual states and tracking errors, and self-adaptive and high-safety robot control is achieved.

The invention discloses a technological process monitoring method and system for a polyester net, and belongs to the technical field of intelligent control, and the method comprises the steps: obtaining real-time technological parameter data in the production process of the polyester net; inputting the real-time process parameter data into a digital twin model, predicting the mesh forming quality of the next period, and generating predicted quality parameters; collecting a polyester mesh surface image, inputting the polyester mesh surface image into the semantic segmentation neural network, identifying a mesh abnormal region, and generating an actual mesh quality score; on the basis of a deviation value between the predicted quality parameter and the actual mesh quality score and energy consumption per unit yield, constructing a reward signal and inputting the reward signal into a deep reinforcement learning controller; the deep reinforcement learning controller outputs a process parameter adjustment instruction and sends the process parameter adjustment instruction to an execution mechanism; and feeding back the actual mesh quality score to the digital twin model, correcting the error of the prediction quality parameter, and updating the model synchronization precision. The product quality stability is improved.

The invention belongs to the technical field of production line monitoring optimization, and particularly relates to an automatic optimization method and system for can-making stamping parameters, and the system comprises a multi-source data collection module which comprises an industrial camera, a material component sensor and an equipment state sensor; the defect intelligent analysis module is used for identifying stamping part defects based on a deep learning model and associating causes; the parameter dynamic optimization module is used for generating an adjusting instruction according to the defect cause; multi-source data collection is matched with defect intelligent analysis, the defect type and the defect position of the can making stamping part are determined, then dynamic weight analysis is carried out, defect causes are matched according to the defect type, then the matched adjusting quantity is obtained, then the adjusting quantity is optimized and verified according to a compensation execution structure and a virtual verification module, and therefore the accuracy of the can making stamping part is improved. The finally obtained adjusting quantity is used for adjusting production line parameters, in addition, closed-loop optimization is carried out through an incremental learning controller, and the system can systematically optimize the problems of slow defect response, high misjudgment rate and poor adjusting precision in the can making stamping field.

The invention provides an underwater robot multi-task rapid adaptive control method based on meta reinforcement learning. The method comprises the following steps: constructing an underwater robot dynamic model and a thrust distribution strategy matrix; three reinforcement learning controllers, namely, a non-overshoot position controller, an overshoot allowing position controller and a propeller flexible control controller, are respectively designed according to diversified task requirements; introducing a meta-learning mechanism to build a meta-training platform, and training the three reinforcement learning controllers to obtain a group of optimal initialization parameters; and deploying the obtained optimal initialization parameters and the subtask reinforcement learning controller to the underwater robot, and carrying out two-stage training according to different tasks. Finally, when the controller is deployed in engineering practice, the output of the propeller can be intelligently adjusted according to the relative distance and speed information, calculated in real time, between the controller and the target position, and it is ensured that accurate position control can be achieved in various task scenes. The method aims at meeting the requirement for rapid self-adaption of multiple tasks of the underwater robot in the complex and changeable underwater environment, and accurate and flexible response to position control is achieved through the control method based on meta reinforcement learning when the task requirements change.

This invention discloses a method for predicting and protecting the overheating risk of asynchronous motors, specifically relating to the field of motor control and protection technology. Based on an intelligent fusion model, it estimates temperature and thermal stress field, calculates the rate of change of thermal stress, non-uniformity, and hotspot trends, and calculates a dynamic health index using historical data. These parameters are then input into a multi-objective reinforcement learning controller to optimize long-term health and short-term performance, generating a thermal shaping control vector to regulate the motor. This invention combines physical mechanisms with data-driven approaches through an intelligent fusion model, utilizing a graph neural network to learn the structure of the heat conduction graph and verify physical laws, thereby improving the accuracy of thermal state estimation. It achieves a multi-dimensional risk characterization combining transient impact and cumulative effects through the rate of change of thermal stress, non-uniformity, hotspot trends, and dynamic health index. By optimizing long-term health and short-term performance losses, a thermal shaping control vector is generated to achieve regulation from passive protection to active prevention, extending the motor's service life.

The present application relates to the technical field of steel plate hot galvanizing production, and discloses an intelligent scheduling system for steel plate hot galvanizing production process.The system comprises a working condition sensing module, which continuously collects the steel strip inlet temperature of the galvanizing pot area, zinc liquid composition concentration fluctuation data, air knife pressure time sequence parameters and roller system operation state signals; a feature extraction module receives the working condition data stream, analyzes the steel strip surface temperature field distribution characteristics and the viscosity variation law of the zinc liquid, and generates a process feature map; a parameter optimization module dynamically adjusts the membership function parameters and rule base weight coefficients of the fuzzy reinforcement learning controller according to the process feature map, and outputs an optimized coating control parameter set; an adaptive coating control module generates a coating thickness control instruction based on the real-time zinc liquid temperature parameters and the optimized coating control rule set; and a dynamic roller system scheduling module realizes global collaborative control of the roller system speed based on the instruction.The system can realize intelligent scheduling of the steel plate hot galvanizing production process.

The application discloses a kind of reinforcement learning batch process control methods based on improved AC algorithm, it is related to the field of deep reinforcement learning and batch process control field.The method will be based on reinforcement learning method The batch process control is modeled as an optimal control problem on the basis of Markov decision process;Control action constraint is introduced in the reward function of reinforcement learning controller, the number of effective reward samples is increased to improve the learning rate of reinforcement learning controller, and the control cycle is shortened.Priority sampling method is introduced in the Actor-Critic algorithm of deep reinforcement learning, and a soft actor-critic algorithm with priority sampling is proposed to improve the sampling efficiency in the experience replay pool.The present application does not depend on prior knowledge and process model, and can realize model-free control of batch process.

The invention relates to a self-adaptive energy management method and device for a fuel cell hybrid electric vehicle and a medium, and the method comprises the steps: collecting vehicle operation parameters in real time, inputting the vehicle operation parameters to a BP neural network driving mode recognizer optimized based on a dung beetle optimization algorithm, and processing the input time sequence operation parameters through a sliding window mechanism, identifying a current driving condition category; adaptively calling a corresponding energy management sub-strategy from a strategy library; each sub-strategy in the sub-strategy library corresponds to a plurality of historical driving condition categories obtained by dividing historical driving conditions through an SMK-means clustering algorithm; each sub-strategy is obtained by training an improved deep reinforcement learning controller which is based on a deep Q network and integrates a noise network, multi-step learning, a double-Q network, a decision network and a priority experience playback mechanism; and calculating and outputting a power distribution instruction between the fuel cell and the power cell. Compared with the prior art, the method has the advantages of high efficiency, high collaboration, high robustness and the like.

The application discloses a circuit system iterative learning control method for processing local Lipshitz nonlinearity under unknown state. The method aims at the actual problem that the circuit system current is unknown, and constructs an adaptive gain state observer based on a reference model. An iterative length selection index is constructed, so that the output of the system and the observer does not violate the given limit range. An iterative learning controller is constructed to control the circuit system with local Lipshitz nonlinearity. Compared with the traditional control method which needs an accurate mathematical model, the control algorithm is more novel and the application condition is simple. When the method is used for repeated task operation, the control precision is effectively improved, and the method has good engineering application value.

The application discloses a safety control method and system for unmanned loader based on machine learning, and relates to the technical field of safety control.The method comprises the following steps: processing multi-source sensor data by using an anti-vibration feature extraction network to generate a safety evaluation coefficient; analyzing a safety state based on a variational autoencoder model to generate structured safety warning information; applying reinforcement learning for multi-objective optimization to output a safe driving trajectory; and finally converting the trajectory into a control instruction through a deep learning controller and adjusting the control parameters in real time based on an online learning mechanism.The application realizes intelligent safety control of the unmanned loader under complex working conditions by constructing a whole-process machine learning processing chain from data acquisition to control execution, and solves the problems of poor safety control adaptability, insufficient cooperation between control modules and vibration interference affecting the accuracy of judgment in the prior art.

A power tool includes a housing and a sensor, a machine learning controller, a motor, and an electronic controller supported by the housing. The sensor is configured to generate sensor data indicative of an operational parameter of the power tool. The electronic controller includes an electronic processor and a memory. The memory includes a machine learning control program. The electronic controller is configured to receive the sensor data. The sensor data includes a motor speed of the motor, a motor current of the motor, and a motion characteristic of the power tool. The electronic controller is configured to process the sensor data using the machine learning control program and generate an output based on the sensor data. The output can include an identified type of application that is being performed by the power tool.

The invention discloses a filament current control method, a filament current control system and an X-ray machine. The filament current control method is used for controlling the filament current of a cathode filament in a bulb tube of the X-ray high-voltage generator, and comprises the following steps: when the bulb tube current of the bulb tube exceeds a preset condition, carrying out iterative learning on the corresponding filament current of each exposure parameter point through an iterative learning controller to obtain an exposure parameter point; and the corresponding filament current preset value IFRefSave of each exposure parameter point is updated after iteration, so that the filament current can be controlled by using the updated corresponding filament current preset value when a user uses each exposure parameter point for exposure again. By adopting the iterative learning method, the problem that the accuracy of short-time exposure is reduced due to the change of the corresponding relationship between the filament current and the bulb tube current caused by the aging reason of the long-time use of the bulb tube in the prior art can be effectively avoided.

The application discloses a mechanical arm iterative learning control method based on a radial basis function network, which solves the limitations of traditional mechanical arm control methods in dealing with nonlinear characteristics and external disturbance problems in a complex dynamic environment. The method comprises the following steps: firstly, constructing a mechanical arm system dynamics model, and adopting a dynamic correction strategy to dynamically correct and optimize a reference trajectory; then, designing a radial basis function neural network to construct a nonlinear compensation term, and designing dynamic weight parameters and dynamic learning gains for optimizing the performance of a controller; finally, designing an iterative learning controller, and verifying the stability and error convergence of the control algorithm. Through the radial basis function network, the dynamic adjustment strategy and the iterative learning strategy, the control precision, adaptability and error convergence speed of the mechanical arm system are improved.

The application discloses a kind of direct current motor iterative learning control optimization methods for executing change task, it is related to direct current motor control field.The method is based on the closed loop feedback control system of direct current motor and parallelly joins iterative learning controller, based on promotion technique will direct current motor control system be converted into time series input-output matrix model.Under the norm optimization framework, optimal iterative learning control algorithm is designed, through the combination of batch-to-batch repeated learning and batch real-time feedback, so that the system basically realizes zero error tracking to expected output.Based on the optimal input sequence and error sequence obtained by repeatedly executing a task, the feedback and feedforward controller are integrated into a new learning-based feedback controller using least squares fitting method, and finally it is applied to the system executing the change trajectory task.This method transfers the historical learning experience to a new task without limiting its time length, and realizes the trajectory tracking of direct current motor change task without relearning.

The invention relates to a reinforcement learning driven self-adaptive multi-mode depth forgery detection method. The method comprises the following steps: constructing a multi-modal data set containing videos and audios, and preprocessing the multi-modal data set; a multi-modal depth forgery detection main model is constructed, the model adopts a parallel isomorphic double-encoder architecture, and each layer of encoder is integrated with a cross-modal fusion module, a frequency domain enhancement module and a contrast learning loss calculation module; constructing a reinforcement learning controller to dynamically regulate and control the on-off of each module; training the multi-modal deep forgery detection model in stages, wherein the stage comprises a main model pre-training stage, a main model and reinforcement learning model combined training stage and a main model fine tuning stage; and finally, inputting a to-be-detected sample into the trained model, and outputting authenticity probability distribution of the to-be-detected sample. According to the method, the problems of calculation redundancy, insufficient cross-modal fusion, frequency domain feature extraction rigidness and the like in an existing multi-modal detection model can be solved, and the detection efficiency, the flexibility and the generalization ability are improved.

The present application belongs to the technical field of photovoltaic module performance test, and relates to a photovoltaic module wind load spectrum simulation test system and method. The system comprises a closed fluid cavity provided with a fluid inlet and a fluid outlet; a fluid pressure control system comprising a pressure storage tank, a proportional valve, an on-off valve and a vacuum pump; the pressure storage tank is connected in sequence through the proportional valve, the on-off valve and the fluid inlet; the air inlet of the vacuum pump is connected with the fluid outlet; a spectrum analysis and control module comprising a data acquisition card and an iterative learning controller; the input end of the data acquisition card is connected with the output end of the pressure sensor, and the output end of the data acquisition card is connected with the input end of the iterative learning controller; and the output end of the iterative learning controller is connected with the proportional valve, the on-off valve and the vacuum pump respectively. The present application can reproduce any target wind pressure spectrum with high precision and fast convergence, and provides a high-fidelity and high-reliability dynamic loading means for wind fatigue resistance and accelerated life test of photovoltaic modules.

The application discloses a kind of event trigger regulation time performance control method based on memory applied to underactuated ASV, including constructing preset regulation time performance function, to obtain trajectory tracking constraint according to initial condition, the function allows to define convergence time, and eliminates the constraint problem that initial error must be within performance boundary;Based on deep neural network DNN, according to ASV dynamics model, obtain the optimization ASV dynamics model containing DNN modeling error, construct modeling error approximation observer, approximate the DNN modeling error in optimization ASV dynamics model, to construct deep learning controller, learn the unknown dynamics of ASV by designing deep learning controller, improve the accuracy of learning and enhance the interpretability of DNN;Based on the memory event trigger mechanism constructed, it is used for optimizing communication resource utilization, and can be dynamically adjusted according to real-time data and historical data, when control signal mutation occurs, memory event trigger mechanism gives priority to historical data, overcome the problem that traditional event trigger mechanism excessively relies on real-time data.

The invention discloses an additive manufacturing process optimization method and system based on jet electrochemistry, and belongs to the technical field of additive manufacturing. The method comprises the steps that the pH value, the ion concentration, the temperature and the current density of a deposition interface micro-area are obtained in real time through a micro-area sensing array integrated in a jet flow nozzle; jet velocity, pressure, electrolyte components and environment interference information are synchronously collected; constructing a uniform process state representation vector for the multi-source heterogeneous data; inputting the vector into a pre-trained multi-physical field coupling process model to realize millisecond-level prediction of the deposition rate, the morphology and the side reaction probability; a reinforcement learning controller is used for generating a self-adaptive adjustment instruction for the voltage, the jet velocity and the nozzle track; the problems that in the prior art, the deposition process is unstable, and real-time micro-area sensing and regulation lag is lacked are solved, in-situ sensing and dynamic regulation of the micron space scale and the millisecond time scale are achieved, and the compactness, the component uniformity and the forming precision of a deposition layer are remarkably improved.