Laser cladding and cutting combined machining system based on machine learning

Abstract

The invention relates to the technical field of laser intelligent manufacturing, and discloses a laser cladding and cutting combined machining system based on machine learning, which comprises three core units, namely a multi-mode sensing core unit, a central processing core unit and a laser machining execution core unit. The multi-mode sensing unit integrates a scanning galvanometer type thermal infrared imager and a non-contact ultrasonic transducer array, and synchronously collects a surface temperature field and internal acoustic impedance. An acoustic-thermal correlation health model is arranged in the central processing unit, and internal hidden defects are identified on line by comparing health acoustic characteristics predicted by real-time temperature with a measured value. Therefore, the feed-forward parameter planning module fuses the physical state field and the defect identification result, dynamically plans and outputs an optimal process parameter sequence, and drives a laser processing execution unit to complete adaptive processing. According to the method, closed-loop self-adaptive regulation and control are realized through acoustic-thermal multi-mode information fusion and feedforward control, formation of internal defects can be effectively predicted and inhibited, and the quality, consistency and reliability of processed finished products are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of laser intelligent manufacturing technology, specifically to a laser cladding and cutting composite processing system based on machine learning. Background Technology

[0002] Laser cladding and cutting, as key material processing technologies, have wide applications in aerospace, energy and power, and mold manufacturing. In conventional processing, process parameters such as laser power, scanning speed, and powder feed rate are usually preset based on experience or offline simulation. This open-loop control method assumes that the processing conditions are stable and uniform. However, in actual processing, the local geometric features of the workpiece, the thermal accumulation effect, and small fluctuations in material properties can all cause dynamic changes in the physical state field, thus affecting the final processing quality.

[0003] To improve processing quality, existing technologies have introduced online monitoring methods. However, these monitoring methods mostly focus on easily observable physical quantities, such as using photodiodes, infrared thermal imagers, or visual cameras to acquire surface information such as the temperature, size, or morphology of the molten pool. These methods cannot directly obtain the internal structural state of the material, and therefore lack effective real-time detection capabilities for internal latent defects that affect the service performance of the final component, such as porosity, cracks, and lack of fusion.

[0004] Therefore, defect detection and assessment often rely on post-processing non-destructive testing methods, such as X-ray or ultrasonic testing. This offline inspection method not only increases production costs and time but also fails to allow real-time intervention in the processing to prevent defects, leading to unstable yield rates. Some existing technologies attempt to use surface information for feedback control, but since there is no stable and direct mapping relationship between surface condition and the formation of internal defects, the control system struggles to make accurate judgments and adjustments. Furthermore, this type of feedback control is inherently lagging, only correcting after deviations occur, and cannot proactively avoid the conditions for defect formation.

[0005] In summary, existing laser processing technologies have significant gaps in their ability to perceive the internal structural state in real time and accurately, and to plan feedforward process parameters based on this to actively suppress internal latent defects. Summary of the Invention

[0006] The technical problem to be solved by this invention is that existing laser cladding and cutting composite processing systems usually use fixed process parameters for open-loop control, which cannot perceive the evolution of the internal physical state of the workpiece and the emergence of latent defects in real time. Therefore, it is difficult to adaptively adjust the processing process, resulting in unstable processing quality and low yield.

[0007] To address the aforementioned technical problems, this invention provides a laser cladding and cutting composite processing system based on machine learning.

[0008] The system includes a multimodal sensing unit, a central processing unit, and a laser processing execution unit. The central processing unit is communicatively connected to the multimodal sensing unit. The laser processing execution unit receives instructions from the central processing unit and executes the operations.

[0009] The multimodal sensing unit is used to simultaneously acquire the surface temperature distribution map and the internal acoustic impedance distribution map of the workpiece to be processed.

[0010] The central processing unit has a built-in physical state field modeling module, an online health status verification module, and a feedforward parameter planning module.

[0011] The physical state field modeling module is used to construct, in real time, a physical state field that characterizes the global physical state of the workpiece to be processed, based on the surface temperature distribution map and the internal acoustic impedance distribution map collected by the multimodal sensing unit.

[0012] The online health status verification module is used to verify the correlation between the surface temperature distribution map and the internal acoustic features extracted from the internal acoustic impedance distribution map based on a preset acoustic-thermal correlation health model, and to generate a latent defect identification result based on the result of the correlation verification.

[0013] The feedforward parameter planning module is used to predict the physical state of future nodes on the processing path based on an enhanced state field that integrates the current physical state field and the latent defect identification results, and a preset processing path, and to generate a spatiotemporally varying process parameter sequence according to the predicted physical state.

[0014] The laser processing execution unit is used to receive the process parameter sequence and perform laser cladding or laser cutting operations according to the process parameter sequence.

[0015] In one specific implementation, the latent defect identification result is an anomaly flag. The enhanced state field is formed by integrating the anomaly flag into the physical state field.

[0016] Preferably, the physical state field modeling module employs a physical information neural network model. The physical information neural network model solves for the physical state field using a loss function that includes a data loss term and a physical law loss term.

[0017] In a specific embodiment, the loss function is constructed by weighted summation of the data loss term and the physical law loss term, and its mathematical expression is as follows:

[0018] L total =λ data L data +λ phys L phys ;

[0019] Among them, L total L represents the value of the total loss function. data L represents the data loss term; phys λ represents the loss term of the physical law; data and λ phys These are the weighting coefficients for the data loss term and the physical law loss term, respectively.

[0020] The data loss item L data This is used to ensure that the output of the physical state field is consistent with the data acquired by the multimodal sensing unit. The physical law loss term L... phys The output of the physical state field is used to constrain the output to satisfy a preset physical law control equation, such as the heat conduction control equation:

[0021]

[0022] Where ρ represents the material density; C p The material's specific heat capacity is represented by T; temperature is represented by t; time is represented by k; and thermal conductivity is represented by k. represents the divergence operator; Q represents the internal heat source term.

[0023] Furthermore, the online health status verification module performs the correlation verification in two steps:

[0024] First, using the acoustic-thermal correlation health model, based on the real-time surface temperature in the surface temperature distribution map, the health acoustic characteristics associated with the real-time surface temperature under normal processing conditions are predicted.

[0025] Then, the healthy acoustic features are compared with the internal acoustic features extracted from the internal acoustic impedance distribution map, and the latent defect identification result is generated based on the comparison result.

[0026] In a preferred embodiment, the online health status verification module compares the health acoustic features with the internal acoustic features by calculating a health index. The health index is obtained by calculating the Mahalanobis distance between the two, and the latent defect identification result is generated based on the health index. The formula for calculating the Mahalanobis distance is:

[0027]

[0028] in:

[0029] D M (x) represents the Mahalanobis distance, the value of which constitutes the health index;

[0030] x represents the internal acoustic feature vector extracted in real time from the internal acoustic impedance distribution map;

[0031] μ represents the health acoustic feature vector predicted by the acoustic-thermal correlation health model;

[0032] S -1 The inverse matrix representing the covariance matrix of the acoustic characteristics under normal processing conditions;

[0033] (·) T This represents the matrix transpose operation.

[0034] In a specific implementation, the feedforward parameter planning module works as follows: when the enhanced state field contains information indicating the existence of latent defects, the feedforward parameter planning module makes the prediction based on the enhanced state field and the processing path, and generates a sequence of process parameters based on the predicted physical state to form a corrective processing strategy.

[0035] As a preferred technical solution, the feedforward parameter planning module generates the process parameter sequence by optimizing a comprehensive cost function. The evaluation terms of the comprehensive cost function include quantitative indicators of the predicted processing quality, process stability, and workpiece health status.

[0036] In one embodiment, the multimodal sensing unit includes a scanning galvanometer infrared thermal imager for acquiring the surface temperature distribution map, and a non-contact ultrasonic transducer array for acquiring the internal acoustic impedance distribution map.

[0037] Furthermore, the sequence of process parameters includes at least two of the following: laser power, scanning speed, and gas pressure, which vary over time.

[0038] This invention provides a laser cladding and cutting composite processing system based on machine learning. It has the following beneficial effects:

[0039] 1. This invention synchronously acquires the surface temperature distribution map and internal acoustic impedance distribution map of the workpiece through a multimodal sensing unit, and combines them with a physical state field modeling module to construct a physical state field characterizing the global physical state of the workpiece in real time. This method fuses surface thermal information with internal acoustic information to obtain a quantitative description of the internal temperature, stress, and phase evolution of the processing area, providing continuous physical state data in both time and space for subsequent defect identification and parameter planning.

[0040] 2. This invention sets up an online health status verification module, which uses an acoustic-thermal correlation health model to predict the health acoustic characteristics based on the real-time surface temperature, and quantitatively compares them with the measured internal acoustic characteristics. This can identify situations where the acoustic response does not match the thermal state due to abnormal internal microstructure of the material, thereby generating hidden defect identification results and realizing online detection of processing defects that are not visible on the surface.

[0041] 3. This invention, through a feedforward parameter planning module, predicts the physical state of future nodes on the processing path based on an enhanced state field containing real-time physical state and latent defect identification results, and generates a spatiotemporally varying process parameter sequence accordingly. This realizes the transformation from fixed open-loop parameter control to feedforward closed-loop control based on physical state prediction, enabling the system to proactively adjust processing parameters before defects occur or expand, forming a corrective processing strategy, thereby improving the uniformity of processing quality. Attached Figure Description

[0042] Figure 1 This is a structural block diagram of a machine learning-based laser cladding and cutting composite processing system according to an embodiment of the present invention.

[0043] Figure 2 This is a flowchart of an online health status verification method according to an embodiment of the present invention;

[0044] Figure 3 This is a flowchart of a feedforward parameter planning method according to an embodiment of the present invention;

[0045] Figure 4 This is a flowchart of a system-wide closed-loop control method according to an embodiment of the present invention. Detailed Implementation

[0046] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0047] See attached document Figure 1 , Figure 1 This is a structural block diagram of a machine learning-based laser cladding and cutting composite processing system according to an embodiment of the present invention. The present invention provides a machine learning-based laser cladding and cutting composite processing system, which may include: a multimodal sensing unit 10, a central processing unit 20, and a laser processing execution unit 30.

[0048] The multimodal sensing unit 10 is used to collect data in real time and synchronously from the processing area of ​​the workpiece to be processed.

[0049] In one specific embodiment, the multimodal sensing unit 10 includes a scanning galvanometer infrared thermal imager 11 for acquiring surface temperature distribution maps, and a non-contact ultrasonic transducer array 12 for acquiring internal acoustic impedance distribution maps. The multimodal sensing unit 10 digitizes the two acquired data streams and transmits them to the central processing unit 20 via a high-speed data interface.

[0050] The central processing unit 20 is the core of the system's computing and decision-making. Its physical entity can be an industrial control computer or an embedded system, including hardware such as a processor and memory.

[0051] The memory stores instructions that can be executed by the processor to implement the following functional modules:

[0052] The system includes a physical state field modeling module 21, an online health status verification module 22, and a feedforward parameter planning module 23.

[0053] The physical state field modeling module 21 has its input end connected to the multimodal sensing unit 10. It is used to receive surface temperature distribution map and internal acoustic impedance distribution map data, and construct a physical state field that characterizes the global physical state of the workpiece.

[0054] The online health status verification module 22, whose input is also connected to the multimodal sensing unit 10, is used to receive surface temperature distribution map and internal acoustic impedance distribution map data, and to perform verification based on a preset model to generate latent defect identification results.

[0055] The feedforward parameter planning module 23 has its inputs connected to the outputs of the physical state field modeling module 21 and the online health status verification module 22, respectively. This module receives the physical state field and the latent defect identification results, and forms an enhanced state field based on these two inputs. Combined with the preset processing path, it finally generates a spatiotemporally varying sequence of process parameters.

[0056] The laser processing execution unit 30 is used to perform specific laser cladding or laser cutting tasks. This unit includes a laser, a scanning galvanometer, a CNC motion platform, and a powder or wire feeding mechanism. Its control input is connected to the output of the central processing unit 20 to receive the process parameter sequence generated by the feedforward parameter planning module 23.

[0057] During system operation, the multimodal sensing unit 10 and the central processing unit 20 are connected through one or more high-speed data communication interfaces, such as Camera Link, Gigabit Ethernet (GigEVision) or CoaXPress interface, to ensure real-time and complete transmission of sensor data.

[0058] The central processing unit 20 and the laser processing execution unit 30 communicate via a real-time industrial bus, such as EtherCAT or Profinet. This connection method ensures that the process parameter sequence generated by the central processing unit 20, which includes parameters such as laser power and scanning speed, can be executed by the laser processing execution unit 30 without delay and with high precision, thereby achieving dynamic control of the processing.

[0059] See attached document Figure 1 The function of the multimodal sensing unit 10 is to provide the central processing unit 20 with real-time, multi-dimensional physical information about the workpiece in the laser-acting area. This unit achieves synchronous, non-contact measurement of the workpiece's surface and internal state through a combination of sensors based on different physical principles.

[0060] In one specific embodiment, the surface temperature distribution map is acquired by a scanning galvanometer infrared thermal imager 11. This imager is integrated with the laser processing head, and its optical axis is coaxial or paraxially parallel to the optical axis of the processing laser beam. Internally, it contains a point or linear array infrared detector and a two-dimensional high-speed scanning galvanometer system. During operation, the central processing unit 20 controls the deflection of the scanning galvanometer system according to the position of the processing laser head, enabling the infrared detector's field of view to accurately and quickly scan the entire laser-material interaction zone and its adjacent heat-affected zone. The thermal imager processes the acquired infrared radiation signal and converts it into temperature values ​​based on calibration data, ultimately outputting a two-dimensional temperature data matrix T(x,y,t) with a timestamp, which is the surface temperature distribution map. Here, x and y are the coordinates of the workpiece surface, and t is the acquisition time. The acquisition frame rate is matched to the processing speed to ensure complete recording of the dynamic behavior of the molten pool.

[0061] The acquisition of the internal acoustic impedance distribution is accomplished by a non-contact ultrasonic transducer array 12. In one embodiment, this array is constructed based on the principle of laser ultrasound. It includes a pulsed excitation laser and a continuous wave detection laser. The short pulse laser emitted by the pulsed excitation laser is focused onto the surface of the workpiece to be processed. The instantaneous absorption of the laser energy causes rapid thermoelastic expansion of the material surface, thereby exciting broadband ultrasonic waves inside the workpiece.

[0062] The ultrasonic wave propagates inside the workpiece and is reflected and scattered when it encounters interfaces with discontinuous acoustic impedance, such as grain boundaries, inclusions, pores, and cracks. A continuous-wave detection laser (e.g., a laser Doppler vibrometer) focuses its beam at a specific detection point on the workpiece surface to receive the minute surface displacements caused by the ultrasonic waves returning from the interior. When the ultrasonic waves reach the surface, they cause instantaneous vibrations in the surface normal, which alter the reflected optical path of the detection laser, thereby generating phase or frequency modulations in the reflected light that can be detected by an interferometer.

[0063] By demodulating the modulation signal of the detection laser, the time-domain waveform signal of the ultrasonic wave at the detection point, i.e., the A-scan signal, can be reconstructed. A non-contact ultrasonic transducer array 12 is used to arrange multiple detection points around the processing area, or the detection points are rapidly moved using a scanning system, thereby acquiring multiple A-scan signals within a region. By analyzing the time-of-flight and amplitude of the echo in each A-scan signal, the acoustic impedance Z at different depths can be calculated. By spatially mapping the acoustic impedances calculated from all acquisition points within a region, the internal acoustic impedance distribution map Z(x,y,z) of that region can be constructed.

[0064] To ensure data consistency, the acquisition operations of the scanning galvanometer infrared thermal imager 11 and the non-contact ultrasonic transducer array 12 are strictly synchronized in time, with the synchronization signal provided uniformly by the central processing unit 20. Spatially, the excitation and detection points of the ultrasonic waves are positioned adjacent to the molten pool, for example, in the solidified area behind the molten pool, to detect the internal condition of the newly formed solidified structure. Through this spatiotemporally synchronized collaborative operation, the multimodal sensing unit 10 can provide the central processing unit 20 with correlated data on surface temperature and internal structural state at the same time and location.

[0065] See attached document Figure 1 The central processing unit 20 processes, analyzes, and makes decisions on the data collected by the multimodal sensing unit 10, and generates control commands. The internal functional logic of this unit is implemented through software programming, and mainly includes a physical state field modeling module 21, an online health status verification module 22, and a feedforward parameter planning module 23.

[0066] The physical state field modeling module 21 receives synchronized surface temperature distribution maps and internal acoustic impedance distribution maps. Its function is to solve for a physical state field that can characterize the evolution of the physical state of the workpiece during processing. This physical state field is a spatiotemporal four-dimensional data structure, specifically composed of a temperature field T(x,y,z,t), a stress field σ(x,y,z,t), and a phase field φ(x,y,z,t), which respectively describe the temperature, stress tensor, and material phase (solid, liquid, gas, etc.) of any point inside the workpiece at any time.

[0067] In one embodiment, the module employs a Physical Information Neural Network (PINN) model to solve for the physical state field. This model takes spatiotemporal coordinates (x, y, z, t) as input and outputs the corresponding physical quantities (T, σ, φ).

[0068] In a specific network architecture implementation, this physical information neural network model can be a fully connected feedforward neural network with multiple hidden layers, also known as a multilayer perceptron. This network consists of an input layer, multiple hidden layers, and an output layer. The input layer receives four-dimensional spatiotemporal coordinates (x, y, z, t), and the output layer outputs multidimensional physical quantities, such as temperature T and the six independent components σ of the stress tensor. ij And the phase indicator function φ. Neurons in the hidden layers employ nonlinear activation functions, such as the hyperbolic tangent or modified linear units. Through automatic differentiation, the partial derivatives of the network's output with respect to the input coordinates of arbitrary order can be precisely calculated; these derivatives are used to construct the physical law loss term L. phys The physical equation residuals in the equations.

[0069] The training or solution process of this model is driven by optimizing a loss function. This loss function L... total Data loss term L data And physical law loss term L phys Weighted composition:

[0070] L total =λ data L data +λ phys L phys ;

[0071] Where, λ data and λ phys These are preset weighting coefficients. Data loss term L data This is used to ensure that the output of the network model remains consistent with the sensor's measured data; for example, by calculating the surface temperature T predicted by the model. pred (x,y,z surface The surface temperature distribution map T obtained by scanning galvanometer infrared thermal imager 11 is compared with that of t). meas The mean square error between (x, y, t). Physical law loss term L. phys The residual form of the known physical governing equations (such as the heat conduction equation, the Navier-Stokes equation, the thermoelastic constitutive equation, etc.) is then embedded into the loss function. For example, for heat conduction, this loss term can be defined as the mean square value of the residual of the heat conduction equation at all points in the solution domain, forcing the model's output to satisfy the law of heat conduction throughout the entire solution domain.

[0072] See attached document Figure 2 , attached Figure 2 This is a flowchart of an online health status verification method according to an embodiment of the present invention. The online health status verification module 22 is used to evaluate the health status of the workpiece in real time and generate latent defect identification results. The core of this module is a pre-trained acoustic-thermal correlation health model. Its specific workflow is shown in the attached figure. Figure 2As shown. This model is built on a large amount of synchronous acoustic and thermal data under normal, defect-free processing conditions, and it describes the deterministic relationship between a specific surface temperature and the expected internal acoustic response.

[0073] The workflow of the online health status verification module 22 is as follows:

[0074] First, the online health status verification module 22 extracts the real-time surface temperature of the current processing position from the surface temperature distribution map obtained by the scanning galvanometer infrared thermal imager 11, and inputs it into the acoustic-thermal correlation health model. Based on this, the model predicts and outputs a healthy acoustic feature vector μ that should exist at that temperature. At the same time, the module extracts the measured internal acoustic feature vector x at the same location from the internal acoustic impedance distribution map obtained by the non-contact ultrasonic transducer array 12.

[0075] The internal acoustic eigenvector x is obtained by calculating and statistically analyzing the internal acoustic impedance distribution map Z(x,y,z) of a specified region.

[0076] In one specific embodiment, the feature vector may include one or more of the following components:

[0077] The acoustic impedance mean, variance, energy of a specific frequency band of energy spectral density within this region, and gray-level co-occurrence matrix (GLCM) features (e.g., contrast, correlation, energy, and homogeneity) used to characterize the internal texture are extracted. By extracting these quantified features, the original acoustic impedance distribution map can be transformed into a numerical vector with fixed dimensions that can characterize its inherent statistical regularities, facilitating subsequent comparison with the healthy acoustic feature vector μ.

[0078] The module then quantifies the degree of deviation between the measured feature x and the health feature μ by calculating the Mahalanobis distance, which constitutes a health index:

[0079]

[0080] Among them, S -1 The inverse of the acoustic feature covariance matrix under healthy conditions is a static parameter determined during the model building phase. Finally, the calculated health index D... M (x) is compared with a preset threshold. If the index exceeds the threshold, it is determined that there is a latent defect at the current location. The preset threshold is determined by statistical analysis of a large number of sample data that are known to be in a healthy state and known to be in a defective state.

[0081] Specifically, a separate validation dataset can be used to calculate the health index of all healthy and defective samples. Through receiver operating characteristic curve analysis, a health index value that achieves an optimal balance between the true positive and true negative rates in defect detection is selected as the preset threshold. This method ensures the statistical reliability of defect determination and generates corresponding latent defect identification results, such as an anomaly flag (with a value of 1) associated with the location coordinates; otherwise, the flag is 0.

[0082] See attached document Figure 3 , attached Figure 3 This is a flowchart of a feedforward parameter planning method according to an embodiment of the present invention. The feedforward parameter planning module 23 is the execution module for implementing adaptive machining control. Its specific workflow is shown in the attached figure. Figure 3 As shown in the diagram, this module first integrates the physical state field output by the physical state field modeling module 21 with the latent defect identification results output by the online health status verification module 22 to form an enhanced state field. This integration process is a data fusion operation; for example, the abnormal flag bit is used as a new data channel and superimposed onto the physical state field data structure of the corresponding coordinates.

[0083] Subsequently, based on this enhanced state field and a preset processing path, the module uses a temporal prediction model (e.g., a Long Short-Term Memory network LSTM or a gated recurrent unit GRU) to predict the physical state of one or more nodes on the processing path in the future.

[0084] Finally, the module generates a spatiotemporally varying sequence of process parameters by optimizing a comprehensive cost function. The goal of this cost function is to optimize the predicted future physical state. Its components may include: terms for evaluating predicted processing quality (e.g., minimizing predicted residual stress, achieving target penetration depth), terms for evaluating process stability (e.g., minimizing predicted melt pool temperature fluctuations), and terms for evaluating workpiece health (e.g., avoiding the emergence of new anomaly flags in the prediction). When the input enhanced state field already contains anomaly flags, the optimization process automatically solves for a set of process parameters that can make the predicted physical state of the defective region tend towards health; this parameter sequence constitutes the remedial processing strategy. The generated sequence of process parameters (e.g., time-varying laser power P(t) and scanning speed v(t)) is sent to the laser processing execution unit 30.

[0085] See attached document Figure 1 The laser processing execution unit 30 is the terminal part that realizes physical processing in the technical solution of the present invention. Its function is to receive and accurately execute the process parameter sequence generated by the central processing unit 20.

[0086] In one specific implementation, the laser processing execution unit 30 may consist of a fiber laser, a laser processing head, a multi-axis CNC motion platform, a material supply system, and a gas control system. The laser processing head integrates a scanning galvanometer for changing the laser beam direction, a focusing lens group for focusing the laser beam, and coaxial or off-axis nozzles for conveying materials and gas. The CNC motion platform is responsible for driving the laser processing head or the workpiece to achieve a preset processing path.

[0087] The internal controller of this unit receives a sequence of process parameters from the central processing unit 20. This sequence is a discrete time-series data stream, which may be in the following form:

[0088] {(t i ,P i ,v i G i ),i=1,2,...,N};

[0089] Among them, t i P represents the i-th time point; i This indicates at time point t i Target laser power; v i This indicates at time point t i Target laser scanning speed; G i This indicates at time point t i The target gas pressure or flow rate.

[0090] After receiving the data stream, the internal controller of this unit parses it and performs analysis at each time point t. i It issues precise control commands to each hardware subsystem. For laser power control, the controller sets the power setpoint P. i It is converted into an analog voltage or digital signal and sent to the power supply of the fiber laser so that at time point t i Output the specified laser power.

[0091] Regarding the control of the scanning speed, the controller sets the scanning speed to v. i This is converted into the angular velocity command required by the scanning galvanometer driver. The scanning galvanometer deflects at a specific angular velocity according to this command, causing the focused laser spot to move at a velocity v on the workpiece surface. i move.

[0092] For gas pressure control, the controller will set the gas pressure or flow rate to G. i It sends signals to the mass flow controller or electronic proportional valve in the gas control system to precisely regulate the supply of protective gas for the cladding process or auxiliary gas for the cutting process.

[0093] The core function of the controller in the laser processing execution unit 30 is to ensure that the control of laser power, scanning speed, gas pressure, and the position of the CNC motion platform is strictly synchronized in time. Through a high-frequency instruction refresh rate and the rapid response of the hardware subsystem, the combination of process parameters planned by the central processing unit 20, which varies in both space and time, is physically and accurately reproduced on the workpiece to be processed, thereby realizing laser cladding or laser cutting operations.

[0094] The following will be combined with the appendix Figure 1 and attached Figure 4 The overall working method and internal information flow of the system of the present invention are described in detail. (Appendix) Figure 4 This is a flowchart of a system-wide closed-loop control method according to an embodiment of the present invention. The flowchart is illustrated in the attached diagram. Figure 1 The system shown illustrates the dynamic information interaction and control loop between its various components.

[0095] Before the laser processing task begins, a preset processing path is loaded into the feedforward parameter planning module 23 of the central processing unit 20. This processing path defines the motion trajectory of the laser processing head relative to the workpiece to be processed.

[0096] When the laser processing execution unit 30 begins to move along the processing path and perform processing, the system enters a continuous closed-loop control process. Within each time step of the process, the multimodal sensing unit 10 starts operating first. Under the synchronous signal control of the central processing unit 20, the scanning galvanometer infrared thermal imager 11 and the non-contact ultrasonic transducer array 12 synchronously acquire data from the current processing point and its adjacent area, obtaining the surface temperature distribution map and the internal acoustic impedance distribution map at that moment, respectively. These two data streams are transmitted to the central processing unit 20 in real time.

[0097] Upon receiving the sensor data, the central processing unit 20 initiates parallel processing through its two internal modules. The physical state field modeling module 21 receives both data streams and solves them using its built-in physical information neural network model, outputting a physical state field characterizing the current global state of the workpiece. Simultaneously, the online health status verification module 22 also receives the same two data streams, performs verification and comparison using its acoustic-thermal correlation health model, and outputs a latent defect identification result characterizing whether a defect exists at the current location.

[0098] Subsequently, the feedforward parameter planning module 23 receives the outputs from the two modules mentioned above, fusing the physical state field and the latent defect identification results to form an enhanced state field containing more complete information. Based on the current enhanced state field and future position information on the processing path, this module predicts the physical state of future nodes. Based on this prediction, it optimizes its internal comprehensive cost function to solve for and generate an optimal, spatiotemporally varying sequence of process parameters. If the enhanced state field contains defect information, the sequence of process parameters generated by this optimization process constitutes a corrective processing strategy.

[0099] The generated sequence of process parameters is sent from the central processing unit 20 to the laser processing execution unit 30. The laser processing execution unit 30 accurately analyzes the sequence and converts it into real-time control commands for hardware such as the laser, scanning galvanometer, and gas system, thereby accurately executing the new processing parameters in the next time step.

[0100] This execution process alters the physical state of the workpiece, and this new state is then acquired by the multimodal sensing unit 10 in the next time step, thus forming a continuous, real-time, and adaptive control closed loop. This closed-loop process continues along the entire machining path until the machining task is completed.

Claims

1. A machine learning-based laser cladding and cutting composite processing system, characterized in that, include: A multimodal sensing unit is used to simultaneously acquire the surface temperature distribution map and the internal acoustic impedance distribution map of the workpiece to be processed; The central processing unit is communicatively connected to the multimodal sensing unit, and the central processing unit has the following built-in features: The physical state field modeling module is used to construct, in real time, a physical state field characterizing the global physical state of the workpiece to be processed, based on the surface temperature distribution map and the internal acoustic impedance distribution map. The online health status verification module is used to verify the correlation between the surface temperature distribution map and the internal acoustic features extracted from the internal acoustic impedance distribution map based on a preset acoustic-thermal correlation health model, and to generate latent defect identification results. The feedforward parameter planning module is used to predict the physical state of future nodes on the processing path based on an enhanced state field that integrates the current physical state field and the latent defect identification results, as well as a preset processing path, and to generate a spatiotemporally varying process parameter sequence according to the predicted physical state. A laser processing execution unit is used to receive the process parameter sequence and perform laser cladding or laser cutting operations according to the process parameter sequence.

2. The laser cladding and cutting composite processing system based on machine learning according to claim 1, characterized in that, The physical state field modeling module adopts a physical information neural network model, which solves the physical state field through a loss function that includes a data loss term and a physical law loss term.

3. The laser cladding and cutting composite processing system based on machine learning according to claim 1, characterized in that, The specific methods by which the online health status verification module performs the correlation verification include: Using the aforementioned acoustic-thermal correlation health model, based on the real-time surface temperature in the surface temperature distribution map, the health acoustic characteristics associated with the real-time surface temperature under normal processing conditions are predicted. The healthy acoustic features are compared with the internal acoustic features extracted from the internal acoustic impedance distribution map to generate the latent defect identification result.

4. The laser cladding and cutting composite processing system based on machine learning according to claim 1, characterized in that, The latent defect identification result is an anomaly flag, and the enhanced state field is formed by integrating the anomaly flag into the physical state field.

5. The laser cladding and cutting composite processing system based on machine learning according to claim 1, characterized in that, The specific working method of the feedforward parameter planning module is as follows: When the enhanced state field contains information indicating the presence of latent defects, the feedforward parameter planning module makes the prediction based on the enhanced state field and the processing path, and generates a sequence of process parameters based on the predicted physical state to form a corrective processing strategy.

6. The laser cladding and cutting composite processing system based on machine learning according to claim 1, characterized in that, The multimodal sensing unit includes: A scanning galvanometer infrared thermal imager used to acquire the surface temperature distribution map; A non-contact ultrasonic transducer array used to acquire the internal acoustic impedance distribution map.

7. The laser cladding and cutting composite processing system based on machine learning according to claim 1, characterized in that, The feedforward parameter planning module generates the process parameter sequence by optimizing a comprehensive cost function, which includes evaluation terms for the predicted processing quality, process stability, and workpiece health status.

8. The laser cladding and cutting composite processing system based on machine learning according to claim 2, characterized in that, The loss function is constructed by weighted summation of the data loss term and the physical law loss term, wherein the physical law loss term is used to constrain the output of the physical state field to satisfy the heat conduction control equation, and the data loss term is used to constrain the output of the physical state field to be consistent with the data collected by the multimodal sensing unit.

9. The laser cladding and cutting composite processing system based on machine learning according to claim 3, characterized in that, The online health status verification module compares the healthy acoustic features with the internal acoustic features by calculating a health index. The health index is used to measure the Mahalanobis distance between the healthy acoustic features and the internal acoustic features. The latent defect identification result is generated based on the health index.

10. The laser cladding and cutting composite processing system based on machine learning according to claim 1, characterized in that, The sequence of process parameters includes at least two of the following: laser power, scanning speed, and gas pressure, which vary over time.

Images