Multi-degree-of-freedom laser melting process monitoring and feedback system

Abstract

The invention provides a multi-degree-of-freedom laser melting process monitoring feedback system, which relates to the field of laser surface treatment, and comprises a multi-degree-of-freedom motion control module, a laser emission module, a visual identification module, a temperature monitoring module and a control module, the multi-degree-of-freedom motion control module comprises a six-degree-of-freedom mechanical arm and a PLC, the laser emission module comprises a laser head, an optical fiber and an optical lens, the temperature monitoring module and the laser head are coaxially installed, an infrared thermometer is adopted, the visual identification module is installed on one side of the laser head, and a high-speed camera is adopted; the control module is electrically connected with the visual identification module, the temperature monitoring module, the laser emission module and the PLC. According to the system, real-time monitoring in the laser melting process is achieved through visual monitoring and temperature feedback, then the laser power and the scanning speed are dynamically adjusted, the laser melting quality is improved, and melting defects are avoided.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of laser surface treatment, and particularly relates to a multi-degree-of-freedom laser melting process monitoring and feedback system. BACKGROUND

[0002] Laser melting is a surface treatment method that uses a high-energy-density laser beam to irradiate the surface of a material, causing the extremely thin area of the material surface layer to rapidly melt in a very short time, and then achieving ultra-fast solidification through an external cooling medium, ultimately forming a modified surface layer that is different from the base in composition, organization and performance. Laser melting can significantly improve the wear and corrosion resistance of materials by changing the microstructure of the material surface layer. During the laser melting process, the key technical parameters that need to be controlled are laser power and scanning speed, which together determine the output energy density of the laser. Too low an energy density cannot achieve the improvement of the wear and corrosion resistance of the base, and too high an energy density will damage the base material, affect the surface state and mechanical properties of the material. Currently, the optimization of the laser melting process is mainly based on single-variable experiments, which analyze the melting effect of the base surface through metallographic characterization to achieve control of the melting process. This method has the following problems: first, it is time-consuming and has a long verification period; second, environmental factors during the test cannot be ignored, resulting in large test errors; and third, repeated material laser melting tests are required, which results in large material loss and high cost.

[0003] Chinese patent document CN115508376A discloses a laser cladding head processing state real-time monitoring system and method. The system can monitor the contamination state of the protective lens of the laser cladding head in real time, accurately identify the lens contamination caused by long-term processing and accidents, and then adjust the protective lens in time to avoid affecting the processing quality or even causing processing failure. At the same time, the system can monitor the powder feeding state of the laser cladding head nozzle in real time. When continuously processing for a long time, especially when continuously processing high-reflectivity materials, the system can accurately identify whether the nozzle is blocked, remind the operator to perform technical processing, and avoid affecting the processing quality or even causing processing failure. The system can monitor the protective lens state and powder feeding state in real time only by relying on the real-time acquisition of coaxial images. However, the system only relies on a visual module to detect the forming state of the laser cladding layer and cannot perform real-time adjustment of the laser output power and scanning speed. After obtaining the forming state of the cladding layer, the output power or scanning speed is adjusted, which has a certain lag, i.e., the adjustment is performed only after pores or cracks appear in the cladding layer, which cannot effectively avoid defects in the laser melting process and affects the final surface treatment effect. SUMMARY

[0004] In view of the problems in the prior art, the purpose of the present application is to provide a multi-degree-of-freedom laser fusion process monitoring feedback system, which realizes real-time monitoring of the surface state of the substrate in the laser fusion process through visual monitoring and temperature feedback, dynamically adjusts the laser power and scanning speed of the laser fusion, improves the quality, uniformity and consistency of the laser fusion, and avoids the generation of fusion defects.

[0005] The purpose of the present application is achieved by the following technical solutions: A multi-degree-of-freedom laser fusion process monitoring feedback system, comprising a multi-degree-of-freedom motion control module, a laser emission module, a visual recognition module, a temperature monitoring module and a control module; the multi-degree-of-freedom motion control module is used for realizing accurate control of the complex motion trajectory of a high-energy beam laser in a three-dimensional space, meeting the matching with the surface of a special-shaped substrate to be processed in the laser fusion process, and ensuring the quality consistency of the laser fusion layer, and comprises a six-degree-of-freedom mechanical arm and a PLC controller, the PLC controller is electrically connected with the six-degree-of-freedom mechanical arm and is used for controlling the motion of the six-degree-of-freedom mechanical arm; the laser emission module is installed at the end of the six-degree-of-freedom mechanical arm and comprises a laser head, an optical fiber and an optical lens, the laser head converts electric energy into a high-energy laser beam and transmits the high-energy laser beam to the surface of a sample through the optical fiber and the optical lens; the temperature monitoring module is coaxially installed with the laser head and adopts an infrared thermometer, the visual recognition module is installed on one side of the laser head and adopts a high-speed camera; the control module is electrically connected with the visual recognition module, the temperature monitoring module, the laser emission module and the PLC controller.

[0006] Based on the further optimization of the above scheme, the six-degree-of-freedom mechanical arm comprises a base, a lifting mechanism, a transverse movement mechanism, a longitudinal movement mechanism, a rotating mechanism and a vibration damping mechanism; the lifting mechanism is arranged on the base and the transverse movement mechanism is arranged on the lifting mechanism, the longitudinal movement mechanism is arranged on the transverse movement mechanism and the rotating mechanism is arranged on the longitudinal movement mechanism; the lifting mechanism, the transverse movement mechanism and the longitudinal movement mechanism are respectively driven by moving servo motors, and the rotating mechanism is driven by a rotating servo motor; the vibration damping mechanism adopts an elastic support structure to reduce the influence of mechanical vibration on the system, and is arranged on the bottom surface of the rotating mechanism; the laser emission module is arranged on the bottom surface of the vibration damping mechanism.

[0007] A real-time monitoring feedback method for a laser fusion process, using the above monitoring feedback system, comprising: Step S1, visual acquisition: acquiring regional images in the laser fusion process in real time through a high-speed camera, and obtaining key features of visual images through image preprocessing, molten pool segmentation and dynamic feature extraction; Step S2, temperature monitoring: fusing a laser heat source model and a Bayesian particle filter to monitor the temperature in the laser fusion process in real time; Step S3, dynamic prediction: predicting the time point after the current parameter adjustment based on the data of visual acquisition and temperature monitoringτ the state (for solving the lag of visual acquisition and temperature monitoring); Step S4, parameter optimization: based on the dynamic model of causal inference, the optimization control of dynamic prediction is realized, so as to complete the accurate adjustment of laser melting parameters (laser power, scanning speed).

[0008] Based on the further optimization of the above scheme, the image preprocessing is specifically: First, the input high-speed camera at time t The laser melting area image I(t,y) , y (y x ,y y ) represents the two-dimensional pixel coordinates; Then, all the pixel points of the image are traversed to obtain the average value And the variance :

[0009] In the formula: represents the set of image pixel points; N represents the number of pixel points; Then, through experimental calibration, the linear proportional relationship between And the variance is obtained:

[0010] In the formula: k represents the empirical coefficient; Finally, for the pixel y = (y x ,y y ) of the pixel position x = (x x ,x y )， The denoised gray value is:

[0011] In the formula: represents the dynamic Gaussian kernel.

[0012] Based on the further optimization of the above scheme, the molten pool segmentation is specifically: First, the image I(t,x,y) is collected by the high-speed camera, and the temperature distribution T(t,x,y) at the same time is collected by the coaxial infrared thermometer (x,y) represents the pixel coordinates, t ​timestamp; and temperature distribution T(t,x,y) bicubic interpolation is performed to make the temperature field consistent with the resolution of the visual image:

[0013] wherein: denotes the bicubic interpolation weight coefficient; Then, the registered temperature field is taken as an additional input channel and spliced with the visual image:

[0014] After that, a U-Net model structure is constructed, including an Encoder, a Decoder and an output layer, wherein the Encoder extracts multiscale features (such as edges, textures and morphologies) of the visual image through convolution (Conv) and pooling (Pooling), the Decoder gradually restores the spatial resolution through upsampling (UpConv) and fusion of the features of the jump connection, and the output layer adopts 1x1 convolution + Sigmoid to output a molten pool probability mask (1 represents a molten pool and 0 represents a background); The loss function is:

[0015] wherein: H , W denotes the height and width of the image; denotes the real mask; T m denotes the melting point of the material; denotes the weight coefficient; denotes the predicted molten pool boundary, which is obtained by performing edge detection on the predicted mask (for example, a Sobel operator is used to calculate the gradient and a threshold is used to screen the boundary pixels):

[0016] wherein: denotes the gradient of the predicted mask; S yz denotes the gradient threshold.

[0017] Based on the further optimization of the above scheme, the dynamic feature extraction is specifically: According to the number of pixels in the segmented region after molten pool segmentation, the molten pool area is obtained A(t) ; and the morphological factor S(t) and the gray uniformity factor G(t) are obtained:

[0018]

[0019] wherein: L(t) represents the circumference of the molten pool; represents the area of the molten pool.

[0020] Based on further optimization of the above scheme, the step S2 is specifically: First, the temperature sequence is collected by the coaxial infrared temperature measuring instrument T r (t) :

[0021] wherein: represents the real measured temperature (i.e. the temperature of the central grid point acted on by the laser); represents the Gaussian white noise, which is calculated according to the temperature collection sequence T r (t) The corresponding spatial position is obtained; and a laser heat source model is constructed by the current laser parameters:

[0022] wherein: Q(t) represents the heat source term; represents the absorption rate of the material to the laser; P(t) represents the laser power at the current t moment; v(t) represents the laser scanning speed at the current t moment; S 0represents the spot area at the current t moment; Then, the temperature measuring area is divided into a two-dimensional grid, and the temperature of each grid point is T i,j (t) The heat conduction differential equation of each grid point is:

[0023] wherein: Q i,j (t) represents the heat source term of the grid point (i,j) ; represents the thermal diffusivity of the material; represents the Laplacian operator;

[0024] wherein: represents the grid step; Finally, the Bayesian particle filter is performed: sequentially through particle initialization, physical model (heat conduction + laser heat source model) driven prediction, weight update, resampling and state estimation, the corrected high-precision temperature sequence is obtained .

[0025] Based on the further optimization of the above scheme, the step S3 is specifically: The input laser parameters are differentiated τ =[ U(t) ] visual feature parameters P(t),v(t) =[ F(t) A(t),S ] temperature parameters Perform stationarity correction:

[0026] The causality between and , and is tested by Granger causality test respectively, and one-way causal pairs are retained and reverse interference is excluded; And construct the cross-correlation function of one-way causal pairs:

[0027] In the formula: Indicates the corresponding lag time of laser parameter change -> vision / temperature; N Indicates the sequence length; Indicates the mean of ; Indicates the mean of ; Similarly, obtain , obtain the lag time when the peak positions of the two are the same , that is, the change of the laser parameter simultaneously triggers the response of the visual feature parameter and the temperature parameter at step; With 3 lag periods, that is, the length of the continuous window is 3 , input the corresponding laser parameters U wi , visual feature parameters F wi and temperature parameters T wi , and introduce the timing information by position coding, and the input after coding is: U en , F en , T en :

[0028] wherein: is the time step position within the window; d represents the encoding dimension (consistent with the Transformer model dimension);

[0029] and the self-attention layer using the Transformer attention mechanism learns the laser parameters (t),G(t) and the visual feature parameters U(t) temperature parameters the temporal correlation of the visual feature parameters F(t)、 and the temperature parameters coupling; using an encoder architecture, future F(t) step visual feature parameters , temperature parameters are generated by autoregressive prediction The model introduces the superheated conduction residual and the solidification velocity residual as physical constraints, and fuses the data fitting loss to generate a total loss function L pr :

[0030]

[0031] wherein: represents the model predicted visual feature vector at future t + s time; represents the real visual feature vector of the lagging visual observation at future t + s time; represents the model predicted temperature feature vector at future t + s time; represents the real temperature feature vector of the lagging visual observation at future t + s time; v s represents the predicted material molten pool solidification velocity; l k represents the material constant; respectively represents the weight coefficient.

[0032] Further optimization based on the above scheme, the self-attention layer using the Transformer attention mechanism learns the laser parameters U,F,T and the visual feature parameters τ temperature parameters Temporal correlation and cross-attention layer strengthen visual feature parameters U(t) Coupling with temperature parameters Specifically, the coupling is: Self-attention: For each modality, i.e. X ={ P(t),v(t)}Learn self-temporal dependence:

[0033] Output:

[0034] In the formula: W Q , W K , W V Indicates a learnable weight; d mo Indicates the model dimension; Cross-attention: Strengthen the physical correlation between visual feature parameters F and temperature parameters T:

[0035] Output:

[0036] In the formula: F en Is the encoded representation of the molten pool feature; T en Is the encoded representation of the temperature field; Force the model to learn the thermal-mechanical coupling law (i.e. the physical mapping of the temperature field to the molten pool shape).

[0037] Based on the further optimization of the above scheme, the step S4 is specifically: First, build a dynamic causal model to quantify the causal effect between laser parameters U(t) [ P(t) ]and ,

[0038] In the formula: Indicates the state transition matrix, which describes the self-evolution influence of the current state on the future state (synchronized with the future state); Indicates the causal effect matrix (the posterior distribution of the causal effect matrix is solved by Bayesian estimation); Indicates the noise term; Then, taking the predicted , As the goal, the current laser parameters are optimized​v(t) Objective function:

[0039] In the formula: F * [ A * ,S * ,G * ] represents the preset optimal visual feature; T * represents the optimal melting temperature of the material; represents a weight coefficient; Constraint condition:

[0040] In the formula: P min represents an effective melting threshold; P max represents a matrix material damage threshold; v min represents the minimum effective scanning speed, v max represents the maximum effective scanning speed; In the cuckoo search algorithm, a causal guidance strategy is added, that is, according to the causal effect matrix , the size of the parameter search step is adjusted (for example: , it is indicated that the causal effect of power on temperature is stronger, and ​ is adjusted preferentially); and the predicted and are substituted into the objective function, and the improved cuckoo search algorithm is used to solve the optimal P (t) , ​ After each iteration, it is checked whether the parameters meet the constraint condition, and if they are out of range, they are projected to the nearest boundary.

[0041] The following are the technical effects of the present application: In the visual recognition process, the adaptive Gaussian filter is used to dynamically suppress plasma noise, effectively solving the contradiction between the traditional fixed kernel "blurred molten pool edge" or "missing noise"; at the same time, in the molten pool segmentation process, the temperature-constrained U-Net model is fused, and the forced segmentation result is combined with the thermodynamic law of the solid-liquid interface, effectively avoiding false molten pool boundaries caused by visual noise, thereby improving the pool area 、 Morphological factor ​With the extraction accuracy of gray uniformity factor, reduce the extraction error. Through the laser heat source model and Bayesian filtering, the heat conduction equation is taken as the state constraint, the temperature extraction accuracy is improved, and the temperature measurement error is reduced. Based on the data of visual acquisition and temperature monitoring, the state of the current parameter adjustment time is effectively predicted ​ , ensure that the prediction window matches the real physical delay, avoid the error caused by prediction ahead or lag; Cooperate with the dynamic model based on causal inference, realize the optimization control of dynamic prediction, and then complete the closed-loop control of "causal precision, forward regulation, stable and reliable", avoid the misjudgment caused by lag.

[0042] The present application significantly shortens the verification period of process optimization experiment through the cooperative matching of visual identification and temperature feedback, ensures the precise control of laser melting process, and improves the quality consistency and uniformity of laser melting treatment, which has significant economic benefit and practical value. BRIEF DESCRIPTION OF DRAWINGS

[0043] ​ The structure block diagram of the monitoring feedback system in the embodiment of the present application. DETAILED DESCRIPTION

[0044] The technical solutions in the embodiments of the present application will be described clearly and completely below. In the following description, specific details such as specific system structures, technologies, etc. are proposed for the purpose of explanation, not for the purpose of limitation, so as to thoroughly understand the embodiments of the present application.

[0045] Embodiment 1: A multi-degree-of-freedom laser melting process monitoring feedback system comprises a multi-degree-of-freedom motion control module, a laser emission module, a visual recognition module, a temperature monitoring module and a control module; the multi-degree-of-freedom motion control module is used for realizing accurate control of a complex motion trajectory of a high-energy beam laser in a three-dimensional space, meeting the matching of a laser melting process with a special-shaped to-be-processed substrate surface and ensuring the quality consistency of a laser melting layer, and comprises a six-degree-of-freedom mechanical arm and a PLC controller, the PLC controller is electrically connected with the six-degree-of-freedom mechanical arm and is used for controlling the motion of the six-degree-of-freedom mechanical arm; the six-degree-of-freedom mechanical arm comprises a base, a lifting mechanism, a transverse movement mechanism, a longitudinal movement mechanism, a rotating mechanism and a vibration damping mechanism; the lifting mechanism is arranged on the base and the transverse movement mechanism is arranged on the lifting mechanism, and the longitudinal movement mechanism is arranged on the transverse movement mechanism and the rotating mechanism is arranged on the longitudinal movement mechanism; the lifting mechanism, the transverse movement mechanism and the longitudinal movement mechanism are respectively driven by moving servo motors, and the rotating mechanism is driven by a rotating servo motor; the vibration damping mechanism adopts an elastic support structure to reduce the influence of mechanical vibration on the system, and is arranged on the bottom surface of the rotating mechanism; the laser emission module is arranged on the bottom surface of the vibration damping mechanism. The laser emission module is installed at the end of the six-degree-of-freedom mechanical arm and comprises a laser head, an optical fiber and an optical lens (the optical lens comprises a collimating mirror, a beam expander, a reflecting mirror and the like), the laser head converts electric energy into a high-energy laser beam, and the high-energy laser beam is transmitted to the surface of a sample through the optical fiber and the optical lens. The temperature monitoring module is coaxially installed with the laser head and adopts an infrared thermometer (the measurement range is 600-2300 DEG C, the measurement accuracy is better than ± 5 DEG C, and the response time is less than or equal to 1 ms), and the visual recognition module is installed on one side of the laser head (i.e. off-axis installation) and adopts a high-speed camera. The control module is electrically connected with the visual recognition module, the temperature monitoring module, the laser emission module and the PLC controller.

[0046] Embodiment 2 As another preferred embodiment of the technical scheme of the present application, a real-time monitoring feedback method for a laser melting process adopts the monitoring feedback system as described in Embodiment 1, and comprises the following steps. Step S1, visual acquisition: a high-speed camera is used to acquire a regional image in a laser melting process in real time, and key features of a visual image are obtained through image preprocessing, molten pool segmentation and dynamic feature extraction; The image preprocessing specifically comprises the following steps: Firstly, an image acquired by the high-speed camera at time t is inputted, and the image is represented as I (x, y), wherein x and y represent two-dimensional pixel coordinates. t Secondly, the image is preprocessed, and the preprocessed image is represented as I (x, y). ​ Thirdly, the image is segmented, and the segmented image is represented as I (x, y). y Fourthly, dynamic features are extracted from the segmented image, and the extracted image is represented as I (x, y). (y x ,y y ) x and y represent two-dimensional pixel coordinates. Afterwards, all pixel points of the image are traversed to obtain the average gray value of the pixel points. With variance :

[0047] In the formula: Represents the set of pixels in an image; N Indicates the number of pixels; Then, through experimental calibration, we obtain Linear proportionality with variance:

[0048] In the formula: k This represents the empirical coefficient (generally 0.3 ≤). k ≤0.8); Ultimately, for pixels y = (y x ,y y ​ x ,x y )， The denoised grayscale value is:

[0049] In the formula: This represents a dynamic Gaussian kernel.

[0050] The specific steps for dividing the molten pool are as follows: First, images are acquired using a high-speed camera. ​ Temperature distribution at the same moment, collected by a coaxial infrared thermometer ​ , ​ Represents pixel coordinates, t For timestamps; and for temperature distribution ​ Bicubic interpolation is performed to make the temperature field consistent with the resolution of the visual image:

[0051] In the formula: Indicates the bicubic interpolation weighting coefficients; Then, the registered temperature field is used as an additional input channel and stitched with the visual image:

[0052] Next, the U-Net model structure is constructed, including an Encoder, a Decoder, and an output layer. The Encoder extracts multi-scale features (such as edges, textures, and shapes) from the visual image through convolution (Conv) and pooling. The Decoder gradually restores the spatial resolution through upsampling (UpConv) and fusing features from skip connections. The output layer uses a 1x1 convolution + Sigmoid function to output a melt pool probability mask. (1 represents the molten pool, 0 represents the background); The loss function is:

[0053] In the formula: H , W Indicates the height and width of the image; Indicates the true mask; T m Indicates the melting point of the material; This represents the weighting coefficient (obtained from experimental data, typically [0.1, 1]). The predicted melt pool boundary is obtained by edge detection of the prediction mask (e.g., calculating gradients using the Sobel operator and thresholding boundary pixels):

[0054] In the formula: This represents the gradient of the predicted mask; S yz This represents the gradient threshold (obtained from experimental data, typically 0.3).

[0055] Dynamic feature extraction specifically involves: The molten pool area is obtained based on the number of pixels in the segmented region after molten pool segmentation. ​ ; And obtain morphological factors ​ With gray uniformity factor ​ :

[0056]

[0057] In the formula: ​ Indicates the perimeter of the molten pool; This indicates the molten pool region.

[0058] Step S2, Temperature Monitoring: Integrating the laser heat source model with Bayesian particle filtering, the temperature during the laser melting and solidification process is monitored in real time; specifically: First, temperature sequences are acquired using a coaxial infrared thermometer. T r(t) :

[0059] where: represents the real measured temperature (i.e. the temperature of the center grid point where the laser acts); represents the Gaussian white noise, according to the temperature acquisition sequence T r (t) the corresponding spatial position is obtained; and a laser heat source model is constructed by the current laser parameters:

[0060] where: ​ represents the heat source term; represents the material absorption rate of the laser (obtained by pre-calibration experiment); ​ represents the laser power at the current t time; ​ represents the laser scanning speed at the current t time; S 0represents the spot area at the current t time; After that, the temperature measurement area is divided into a two-dimensional grid, and the temperature of each grid point is T i,j (t) The heat conduction differential equation of each grid point is:

[0061] where: Q i,j (t) represents the heat source term of the grid point ​ ; represents the material thermal diffusivity (obtained by experimental measurement); represents the Laplace operator;

[0062] where: represents the grid step; Finally, Bayesian particle filtering is performed: sequentially through particle initialization, physical model (heat conduction + laser heat source model) driven prediction, weight update, resampling and state estimation, Particle initialization: N particles are generated, each particle is a vector of grid point temperatures (such as ); the initial measurement value is T r (0) ; Physical model driven prediction: for each particlek Current state , extract temperature at each grid point T i,j,k (t) ; substitute into discretized state equation (e.g. Euler method or Runge-Kutta method), numerically solve , obtain predicted state ; Weight update: obtain particle weight according to measurement equation, the smaller the difference, the higher the weight:

[0063] In the formula: T pr,k (t) denotes the temperature at the measurement position in the predicted value of the particle k ; denotes the weight of the k th particle; denotes the "proportional to" symbol; denotes the measurement noise variance (obtained by prior test or statistical estimation); After weight normalization, it satisfies:

[0064] Resampling: if the effective particle number is too low (for example: ), the high-weight particles are retained by resampling: sampling randomly according to weight , copying high-weight particles and eliminating low-weight particles to generate a new particle set ; State estimation: obtain the optimal temperature estimation by weighted average:

[0065] Repeat the physical model driven prediction-weight update-resampling-state estimation step by step over time, update the particle state and estimate the temperature, and finally output the corrected high-precision temperature sequence .

[0066] Step S3, dynamic prediction: based on the data of visual acquisition and temperature monitoring, predict the state at the time ​ after the current parameter adjustment (to solve the hysteresis of visual acquisition and temperature monitoring); specifically: adopt difference method to correct the input laser parameters ​ [ ​ ], visual feature parameters ​ [ ​ ​ ], and temperature parameters :

[0067] The causality between the laser parameter change and the visual / temperature response is verified by Granger causality test. The causality between the laser parameter change and the visual / temperature response is verified by Granger causality test. , The causality between the laser parameter change and the visual / temperature response is verified by Granger causality test. The causality between the laser parameter change and the visual / temperature response is verified by Granger causality test. The mutual correlation function of the one-way causal pair is constructed:

[0068] In the formula: represents the corresponding lag time of the laser parameter change→visual / temperature response; N represents the sequence length; represents the mean of ; represents the mean of ; Similarly, obtain , and obtain the lag time when the peak positions of the two are the same , that is, the change of the laser parameter simultaneously triggers the response of the visual feature parameter and the temperature parameter at steps later; With a lag period of 3, that is, a continuous window length of 3 , the corresponding laser parameter U wi , visual feature parameter F wi and temperature parameter T wi are input, and the time sequence information is introduced by position coding. After coding, the input is: U en , F en , T en :

[0069] In the formula: is the position of the time step in the window; d represents the coding dimension (consistent with the dimension of the Transformer model);

[0070] The self-attention layer of the Transformer attention mechanism is used to learn the time sequence correlation of the laser parameter ​ and the visual feature parameter ​ and the temperature parameter , and the cross-attention layer is used to strengthen the coupling of the visual feature parameter ​ and the temperature parameter . Specifically: Self-attention: For each modality, i.e. X ={ ​} learns the self- temporal dependency:

[0071] Output:

[0072] where W Q 、 W K 、 W V denote learnable weights (obtained through experimental data); d mo denote model dimensions; Cross-attention: strengthens the physical correlation between visual feature parameters F and temperature parameters T:

[0073] Output:

[0074] where F en is the encoded representation of the molten pool features; T en is the encoded representation of the temperature field; forces the model to learn the thermal-mechanical coupling law (i.e., the physical mapping from the temperature field to the molten pool morphology); uses an encoder architecture to generate future ​ step visual feature parameters and temperature parameters through autoregressive prediction; The model introduces an overheat conduction residual and a solidification speed residual as physical constraints, and fuses a data fitting loss to generate a total loss function L pr :

[0075]

[0076] where denotes the model-predicted future t + s time visual feature vector; denotes the true visual feature vector of the future t + s time lagged visual observation; denotes the model-predicted future t + s time temperature feature vector; real temperature feature vector representing the future time instant of the hysteresis visual observation t s ) time instant; v s representing the predicted material molten pool solidification speed; l k representing the material constant; representing the weight coefficient (obtained through experimental data), respectively.

[0077] Step S4, parameter optimization: based on the dynamic model of causal inference, the optimization control of dynamic prediction is realized, so as to complete the precise adjustment of the laser melting parameters (laser power, scanning speed), specifically: Firstly, the dynamic causal model is constructed, and the causal effect between the laser parameters ​ [ ​ ] and , is quantified:

[0078] In the formula: represents the state transition matrix, which describes the self-evolution influence of the current state on the future state (synchronized with the future state acquisition); represents the causal effect matrix (the posterior distribution of the causal effect matrix is solved through Bayesian estimation); represents the noise term (obtained through model fitting residual estimation); After that, the predicted , is taken as the target, and the current laser parameter ​ is optimized, and the objective function is:

[0079] In the formula: F * [ A * ,S * ,G * represents the preset optimal visual feature; T * represents the optimal melting temperature of the material (obtained through prior calibration); represents the weight coefficient (obtained through empirical data and experimental data); Constraint condition:

[0080] In the formula: P min represents the effective melting threshold value; P max ​represents the threshold of the base material damage; v min represents the minimum effective scanning speed, v max represents the maximum effective scanning speed; In the cuckoo search algorithm, a causal guidance strategy is added, that is, according to the causal effect matrix , the size of the search step of the adjustment parameter (for example: , it is indicated that the causal effect of power on temperature is stronger, and is preferentially adjusted ​ ); the predicted in step S3 is substituted into the objective function, and the improved cuckoo search algorithm is used to solve the optimal P , (t) , ​ After each iteration, it is checked whether the parameters meet the constraint conditions, and if they are out of bounds, they are projected to the nearest boundary.

Claims

1. A multi-degree of freedom laser fusion process monitoring feedback system characterized by: The application relates to a laser cladding system, comprising a multi-degree-of-freedom motion control module, a laser emission module, a visual recognition module, a temperature monitoring module and a control module; the multi-degree-of-freedom motion control module comprises a six-degree-of-freedom mechanical arm and a PLC controller, the PLC controller is electrically connected with the six-degree-of-freedom mechanical arm and is used for controlling the motion of the six-degree-of-freedom mechanical arm; the laser emission module is installed at the end of the six-degree-of-freedom mechanical arm and comprises a laser head, an optical fiber and an optical lens, the laser head converts electric energy into a high-energy laser beam, and the high-energy laser beam is transmitted to the surface of a sample through the optical fiber and the optical lens; the temperature monitoring module is coaxially installed with the laser head and adopts an infrared thermometer, the visual recognition module is installed on one side of the laser head and adopts a high-speed camera, and the control module is electrically connected with the visual recognition module, the temperature monitoring module, the laser emission module and the PLC controller.

2. A real-time monitoring feedback method for a multi-degree-of-freedom laser fusion process monitoring feedback system according to claim 1, characterized in that: The application relates to a laser cladding system, comprising a multi-degree-of-freedom motion control module, a laser emission module, a visual recognition module, a temperature monitoring module and a control module; the multi-degree-of-freedom motion control module comprises a six-degree-of-freedom mechanical arm and a PLC controller, the PLC controller is electrically connected with the six-degree-of-freedom mechanical arm and is used for controlling the motion of the six-degree-of-freedom mechanical arm; the laser emission module is installed at the end of the six-degree-of-freedom mechanical arm and comprises a laser head, an optical fiber and an optical lens, the laser head converts electric energy into a high-energy laser beam, and the high-energy laser beam is transmitted to the surface of a sample through the optical fiber and the optical lens; the temperature monitoring module is coaxially installed with the laser head and adopts an infrared thermometer, the visual recognition module is installed on one side of the laser head and adopts a high-speed camera, and the control module is electrically connected with the visual recognition module, the temperature monitoring module, the laser emission module and the PLC controller. The application relates to a laser cladding system, comprising a multi-degree-of-freedom motion control module, a laser emission module, a visual recognition module, a temperature monitoring module and a control module; the multi-degree-of-freedom motion control module comprises a six-degree-of-freedom mechanical arm and a PLC controller, the PLC controller is electrically connected with the six-degree-of-freedom mechanical arm and is used for controlling the motion of the six-degree-of-freedom mechanical arm; the laser emission module is installed at the end of the six-degree-of-freedom mechanical arm and comprises a laser head, an optical fiber and an optical lens, the laser head converts electric energy into a high-energy laser beam, and the high-energy laser beam is transmitted to the surface of a sample through the optical fiber and the optical lens; the temperature monitoring module is coaxially installed with the laser head and adopts an infrared thermometer, the visual recognition module is installed on one side of the laser head and adopts a high-speed camera, and the control module is electrically connected with the visual recognition module, the temperature monitoring module, the laser emission module and the PLC controller. The image preprocessing is specifically as follows: Step S3, dynamic prediction: based on the data of visual acquisition and temperature monitoring, predict the state at the time after the current parameter adjustment I(t,y) ; ) of the neighboring pixel position x= (x 3. A method of real-time monitoring feedback for a multi-degree-of-freedom laser fusion process according to claim 2, characterized in that: The molten pool segmentation is specifically as follows: First, the input high-speed camera at time t laser-fused region image I(t,x,y) , y = (y x ,y y ) denotes a two-dimensional pixel coordinate;​ After that, traverse all the pixel points of the image to obtain the average value of its gray scale and variance : In the formula: represents a set of image pixel points; N represents the number of pixel points; Then, by experimental calibration, we obtain Linear proportionality to the variance: In the formula: k represents an empirical coefficient; Finally, for the pixel y = (y x ,y y T(t,x,y) x ,x y )， The de-noised gray value is: In the formulae: denotes a dynamic Gaussian kernel.

4. A method of real-time monitoring feedback for a multi-degree-of-freedom laser fusion process according to claim 2 or 3, characterized in that: (x,y) First, images are captured by a high-speed camera T(t,x,y) Temperature distribution at the same time captured by coaxial infrared thermometer Then, the registered temperature field is taken as an additional input channel and is spliced with the visual image: , The loss function is as follows: represents the pixel coordinates, t is the timestamp; and the temperature distribution The dynamic feature extraction is specifically as follows: is bicubic interpolated so that the temperature field is consistent with the visual image resolution: wherein: denotes the bi-cubic interpolation weight coefficients; A(t) Then, the U-Net model structure is constructed, including an Encoder, a Decoder and an output layer, wherein: the Encoder extracts multi-scale features of the visual image through convolution and pooling; the Decoder gradually restores the spatial resolution through upsampling and fusing the features of the jump connection, and the output layer adopts 1x1 convolution + Sigmoid to output the molten pool probability mask ; S(t) In the formula: H , W represents the height, width of the image; represents the real mask; T m represents the melting point of the material; represents the weight coefficient; representing a predicted molten pool boundary, obtained by edge detection on the prediction mask: In the formula: denotes a gradient of the prediction mask; S yz denotes a gradient threshold value.

5. The method of real-time monitoring and feedback for multi-degree-of-freedom laser fusion process according to claim 2 or 4, characterized in that: G(t) According to the pixel number of the divided area after the molten pool is divided, the molten pool area is obtained L(t) ; and obtain a morphology factor The step S2 is specifically as follows: with a gray level uniformity factor And a laser heat source model is constructed through the current laser parameters: : wherein: Q(t) represents the circumference of the molten pool; represents the area of the molten pool.

6. A method of real-time monitoring feedback for a multi-degree-of-freedom laser fusion process according to claim 5, characterized in that: P(t) Firstly, the temperature sequence is collected by coaxial infrared thermometer T r (t) : In the formula: represents the real measured temperature; represents a Gaussian white noise, according to the temperature acquisition sequence T r (t) corresponding spatial position is obtained; v(t) In the formula: (i,j) represents a heat source term; represents the absorption rate of the material to the laser; The step S3 is specifically as follows: represents the laser power at the current t time; U(t) represents the laser scanning speed at the current t time; S 0represents the spot area at the current t time; After that, the temperature measuring area is divided into two-dimensional grid, and the temperature of each grid point is T i,j (t) The heat conduction differential equation of each grid point is: wherein: Q i,j (t) represents a heat source term at the grid point P(t),v(t) represents the thermal diffusivity of the material; represents the Laplacian operator;​ In the formulae: denotes the grid step size; Finally, the Bayesian particle filter is performed: sequentially through particle initialization, physical model driven prediction, weight update, resampling and state estimation, the corrected high-precision temperature sequence is obtained .

7. A method of real-time monitoring feedback for a multi-degree-of-freedom laser fusion process according to claim 6, characterized in that: F(t) using a difference method on input laser parameters A(t),S(t),G [ And a mutual correlation function of the one-way causal pair is constructed: ] visual feature parameters U(t) [ F(t) (t) ] temperature parameters performing stationarity correction: The Granger causality test is used to test the causality between and , and respectively, and the one-way causal pair is reserved, and the reverse interference is discharged. τ In the formulae: represents the hysteresis time of the laser parameter change -> visual / temperature correspondence; N represents the sequence length; represents the mean value of ; represents the mean value of ; represents the mean value of ; represents the mean value of The same is obtained , the time lag is obtained when the peak positions of both are the same , i.e. the change in the laser parameter simultaneously triggers the response of the visual feature parameter and the temperature parameter at the next step; With 3 lag periods, i.e. the length of the continuous window is 3 corresponding laser parameters U wi , visual feature parameters F wi and temperature parameters T wi , and through position coding to introduce timing information, the input after coding is: U en , F en , T en : In the formulae: is the time step position within the window; d denotes the encoding dimension; and a self-attention layer using the Transformer attention mechanism learns laser parameters The step S4 is specifically as follows: coupled with visual feature parameters F (t)、 temperature parameters and cross-attention layers reinforce visual feature parameters U(t) coupled with temperature parameters using an encoder architecture to generate future P(t),v(t) steps of visual feature parameters , temperature parameters ; The model introduces superheat conduction residual and solidification velocity residual as physical constraints, fuses data fitting loss to generate a total loss function L pr : wherein: represents a model predicted future (t + T) visual feature vector; t s represents a lagged visual observed future (t + T) visual feature vector; t s represents a model predicted future (t + T) temperature feature vector; t s represents a lagged visual observed future (t + T) temperature feature vector; t s v s represents a predicted material weld pool solidification velocity; l k represents a material constant; represents a weight coefficient, respectively.​​​​​​​​ 8. The method of real-time monitoring feedback for multi-degree-of-freedom laser fusion process according to claim 7, characterized in that: U(t) First, a dynamic causal model was constructed to quantify the causal effects of laser parameters The constraint condition is as follows: [ P(t) ] and , on the response. wherein: denotes the state transition matrix, describing the self-evolution influence of the current state on the future state; denotes the causal effect matrix; denotes the noise term; Afterwards, the predicted , current laser parameters v(t) are optimized, the objective function: In the formula: F * [ A * ,S * ,G * ] represents a preset optimal visual feature; T * represents the optimal melting temperature of the material; represents a weight coefficient; ​ wherein: P min represents the effective dissolution threshold; P max represents the matrix material damage threshold; v min represents the minimum effective scan speed, v max represents the maximum effective scan speed; A causal guidance strategy is incorporated into the cuckoo search algorithm, namely, based on the causal effect matrix. Adjust the size of the search step size parameter; use the predicted value obtained in step S3. and Substitute the values ​​into the objective function and use the improved cuckoo search algorithm to solve for the optimal value. ​ , ​ After each iteration, check whether the parameters meet the constraints. If they exceed the limits, project them to the nearest boundary.

Images