Robot control method and system for aluminum material surface spraying treatment

Abstract

The invention relates to the technical field of spraying treatment, and provides a robot control method and system for spraying treatment of the surface of an aluminum material, and the method comprises the following steps: obtaining multispectral information of the surface of the aluminum material, and extracting a multispectral feature value according to the multispectral information; obtaining multi-source noise interference information, and obtaining a noise correction factor according to the multi-source noise interference information; constructing a defect feature matrix according to the multispectral feature value and the noise correction factor, performing path planning according to the defect feature matrix, and generating a control instruction according to the planned path; and the robot is controlled to act according to the control instruction, and spraying treatment is conducted on the surface of the aluminum material. According to the method, path planning is carried out according to the defect feature matrix, the influence of noise of different dimensions on the aluminum material surface defect recognition accuracy is considered, the aluminum material surface defect recognition accuracy can be improved, and the spraying quality is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of spraying treatment, and specifically, to a robot control method and control system for spraying treatment on the surface of aluminum materials. Background Art

[0002] Before leaving the factory, it is generally necessary to spray the surface of aluminum materials to enhance their corrosion resistance.

[0003] After retrieval, some typical prior arts are found. For example, an aluminum profile surface spraying device and its spraying method with the application number CN202010990323.5 can prevent the remaining paint in the material cavity, avoid waste of paint, and prevent the paint from condensing into solid particles in the material cavity, blocking the injection holes during the next spraying operation, and also avoid various problems such as difficult adjustment of the spray pattern and uneven spraying, ensuring the spraying effect and quality of the aluminum profile surface. Another example is an aluminum profile spraying method with the application number CN201911148656.7, the sprayed aluminum profile has a decorative effect of gradient color and low spraying cost. Still another example is an electrostatic spraying automatic control method for aluminum profiles with the application number CN202110126716.6. The operator only needs to perform simulation and adjustment according to the required coating thickness of the aluminum profile to obtain the optimal spraying parameters and speed parameters, and input the optimal spraying parameters and adjust to the optimal speed on the industrial control computer software to obtain an aluminum profile with uniform thickness, realizing the automatic control of electrostatic spraying of aluminum profiles.

[0004] Therefore, for the spraying on the surface of aluminum materials, there are still many technical problems to be solved urgently in actual applications and many solutions have not been proposed. Summary of the Invention

[0005] Based on this, in order to improve the spraying quality of the aluminum material surface, the present invention provides a robot control method and control system for spraying treatment on the surface of aluminum materials. The specific technical solutions are as follows:

[0006] A robot control method for spraying treatment on the surface of aluminum materials, which includes the following steps:

[0007] Obtain the multispectral information of the aluminum material surface, and extract the multispectral feature values according to the multispectral information;

[0008] Obtain the multi-source noise interference information, and obtain the noise correction factor according to the multi-source noise interference information;

[0009] Construct a defect feature matrix according to the multispectral feature values and the noise correction factor, perform path planning according to the defect feature matrix, and generate a control instruction according to the planned path;

[0010] Control the robot action according to the control instruction to perform spraying treatment on the aluminum material surface.

[0011] The robot control method obtains a noise correction factor, constructs a defect feature matrix based on the multispectral eigenvalue and the noise correction factor, and performs path planning according to the defect feature matrix. It considers the influence of different-dimensional noises on the recognition accuracy of aluminum surface defects and can improve the recognition accuracy of aluminum surface defects. On this basis, the aluminum surface defect information is recognized based on the defect feature matrix, and then path planning is performed, which can better spray the aluminum surface and thus improve the spraying quality.

[0012] Preferably, the specific method for obtaining the noise correction factor according to the multi-source noise interference information includes:

[0013] Obtain the noise correction factor at the target position at time t-1, and obtain the time-domain recursion term according to the noise correction factor;

[0014] Perform multi-scale noise decomposition on the defect feature matrix at the target position, and obtain the multi-scale decomposition term according to the decomposition result;

[0015] Obtain the defect confidence at the target position, and obtain the defect confidence weight term according to the defect confidence;

[0016] Obtain the environmental disturbance vector at time t, and obtain the environmental disturbance compensation term according to the environmental disturbance vector;

[0017] Obtain the noise correction factor at time t according to the time-domain recursion term, the multi-scale decomposition term, the confidence weight term, and the environmental disturbance compensation term.

[0018] Preferably, the specific method for performing path planning according to the defect feature matrix includes:

[0019] Extract the spatial feature of the defect feature matrix through CNN and extract the temporal feature of the defect feature matrix through LSTM;

[0020] Fuse the spatial feature and the temporal feature to obtain the defect probability at the target position;

[0021] Obtain the trajectory tracking error term according to the actual path and the reference path, obtain the defect gradient field according to the defect probability, and obtain the defect avoidance term according to the defect gradient field;

[0022] Obtain the path optimization objective function according to the trajectory tracking error term and the defect avoidance term, and perform path planning according to the path optimization objective function.

[0023] Preferably, the specific method for generating a control instruction according to the planned path includes:

[0024] Construct an RBFNN model, train the RBFNN model, and obtain the weight matrix;

[0025] Using the environmental disturbance vector, defect probability, and path optimization objective function as the inputs of the RBFNN model for high-dimensional space mapping;

[0026] Construct a PID controller and obtain the PID control error according to the PID controller;

[0027] Obtain the comprehensive spraying parameter vector based on the RBFNN model and the PID control error, and generate a control instruction according to the comprehensive spraying parameter vector.

[0028] Preferably, the noise correction factor ε(t,x,y) at time t = TD + MRA·CW + EDC;

[0029] where TD = ρ·ε t-1 represents the time-domain recursion term, represents the scale decomposition term, represents the confidence weight term, represents the environmental disturbance compensation term, ρ represents the time decay factor, ε t-1 represents the noise correction factor at time t - 1 for the target position, ω k represents the weight coefficient of the k-th scale decomposition, D k (x,y) represents the defect feature matrix at the target position (x,y), Ψ k (D k (x,y)) represents the decomposition result of the defect feature matrix at the target position at the k-th scale. K represents the number of frequency bands of the scale decomposition, C d represents the defect confidence at the target position, e represents the natural constant, T represents the defect confidence activation threshold, η represents the defect confidence sensitivity coefficient, γ represents the environmental disturbance compensation coefficient, E env (t) represents the environmental disturbance vector, σ represents the normalization parameter.

[0030] Preferably, the path optimization objective function

[0031] where, respectively represent the trajectory tracking error term and the defect avoidance term, P ref 、P act respectively represent the reference path and the actual path, ||·|| represents the Euclidean distance norm, α represents the trajectory tracking weight coefficient, M d represents the defect probability, represents the defect gradient field of the defect probability in the x direction, β represents the defect avoidance weight coefficient, t0 and t1 respectively represent the start and end times of the spraying operation, C represents the integration path.

[0032] Preferably, according to the formula f = RBFNN(W T ·[Md , J path , E env ) + PID(Δe) to obtain the comprehensive spraying vector;

[0033] Among them, RBFNN(W T ·[M d , J path , E env ), PID(Δe) respectively represent the RBFNN model and the PID control error, W represents the weight matrix, and W T represents the transpose of the weight matrix.

[0034] A robot control system for surface spraying treatment of aluminum materials, used to implement the described robot control method, which includes:

[0035] A multispectral module, used to obtain the multispectral information of the aluminum material surface and extract multispectral feature values according to the multispectral information;

[0036] A correction factor acquisition module, used to obtain multi-source noise interference information and obtain a noise correction factor according to the multi-source noise interference information;

[0037] A path planning module, used to construct a defect feature matrix according to the multispectral feature values and the noise correction factor, perform path planning according to the defect feature matrix, and generate control instructions according to the planned path;

[0038] A robot control module, which controls the robot's actions according to the control instructions and performs spraying treatment on the aluminum material surface.

[0039] Preferably, the correction factor acquisition module includes:

[0040] A time-domain recursive term acquisition unit, used to obtain the noise correction factor at the target position at time t - 1 and obtain the time-domain recursive term according to the noise correction factor;

[0041] A multi-scale decomposition unit, used to perform multi-scale noise decomposition on the defect feature matrix at the target position and obtain the multi-scale decomposition term according to the decomposition result;

[0042] A defect confidence acquisition unit, used to obtain the defect confidence at the target position and obtain the defect confidence weight term according to the defect confidence;

[0043] An environmental disturbance acquisition unit, used to obtain the environmental disturbance vector at time t and obtain the environmental disturbance compensation term according to the environmental disturbance vector;

[0044] A correction factor acquisition unit, used to obtain the noise correction factor at time t according to the time-domain recursive term, the multi-scale decomposition term, the confidence weight term, and the environmental disturbance compensation term.

[0045] Preferably, the path planning module includes:

[0046] A feature extraction unit for extracting the spatial features of the defect feature matrix through CNN and the temporal features of the defect feature matrix through LSTM;

[0047] A feature fusion unit for fusing the spatial features and the temporal features to obtain the defect probability at the target position;

[0048] An objective function acquisition unit for obtaining a trajectory tracking error term according to the actual path and the reference path, obtaining a defect gradient field according to the defect probability, and obtaining a defect avoidance term according to the defect gradient field;

[0049] A path planning unit for obtaining a path optimization objective function according to the trajectory tracking error term and the defect avoidance term, and performing path planning according to the path optimization objective function. Description of the Drawings

[0050] The present invention can be further understood from the following description in conjunction with the drawings. The components in the drawings are not necessarily drawn to scale, but the emphasis is placed on showing the principles of the embodiments. In different views, the same reference numerals designate corresponding parts.

[0051] Figure 1 is a schematic diagram of the overall process of the robot control method in an embodiment of the present invention;

[0052] Figure 2 is a schematic diagram of the specific method for obtaining the noise correction factor in an embodiment of the present invention;

[0053] Figure 3 is a schematic diagram of the specific method for performing path planning in an embodiment of the present invention;

[0054] Figure 4 is a schematic diagram of the specific method for generating a control instruction in an embodiment of the present invention.

[0055] Figure 5 is a schematic diagram of the overall structure of the robot control system in an embodiment of the present invention;

[0056] Figure 6 is a schematic diagram of the functional module of the correction factor acquisition module in an embodiment of the present invention;

[0057] Figure 7 is a schematic diagram of the functional module of the path planning module in an embodiment of the present invention;

[0058] Figure 8 is a schematic diagram of the functional module of the robot control system in another embodiment of the present invention Figure 1 ;

[0059] Figure 9 This is a schematic diagram of the functional modules of the robot control system in another embodiment of the present invention Figure 2 . Specific embodiments

[0060] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below in conjunction with its embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and do not limit the protection scope of the present invention

[0061] It should be noted that when an element is referred to as being "fixed to" another element, it can be directly on the other element or there can also be a central element. When an element is considered to be "connected" to another element, it can be directly connected to the other element or there may be a central element at the same time. The terms "vertical", "horizontal", "left", "right" and similar expressions used herein are only for the purpose of illustration and do not represent the only implementation

[0062] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the technical field to which the present invention belongs. The terms used in the description of the present invention herein are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "and / or" used herein includes any and all combinations of one or more of the related listed items

[0063] The "first" and "second" in the present invention do not represent specific quantities and orders, but are only used for name distinction

[0064] Before specifically elaborating on the embodiments of the present invention, a brief introduction to the prior art will be given

[0065] With the development of robot technology, more and more factories choose to use robots to spray the surface of aluminum materials to improve the spraying efficiency and quality of the aluminum material surface. For the robot control method applied to the spraying treatment of the aluminum material surface, generally, the spraying path will be planned first, and the corresponding spraying parameters will be set according to different spraying processes

[0066] There are often more or less defects on the surface of aluminum materials, such as pits, bubbles, scratches and deformations. The types and degrees of the defects will affect the final spraying quality. However, in the existing robot control methods applied to the spraying treatment of the aluminum material surface, when planning the spraying path, the defects on the aluminum material surface are often ignored, and the spraying path is not planned based on the defects on the aluminum material surface, and the spraying quality needs to be further improved

[0067] In order to improve the spraying quality of the aluminum material surface, an embodiment of the present invention provides a robot control method for spraying treatment of the aluminum material surface, asFigure 1 As shown, it includes the following steps:

[0068] S1. Obtain the multispectral information of the aluminum material surface, and extract the multispectral feature values according to the multispectral information.

[0069] The multispectral information can be obtained by a multispectral sensor and a high-precision industrial camera (such as a 3D line array scanning instrument), and it includes multimodal information such as the surface texture, depth, and spectral reflectance of the aluminum material. Specifically, the surface of the aluminum material can be synchronously scanned by the multispectral sensor to generate raw images of multiple channels. At the same time, the spatial alignment of the multispectral raw images can be achieved through calibration plate correction to eliminate geometric distortion.

[0070] S2. Obtain the multi-source noise interference information, and obtain the noise correction factor according to the multi-source noise interference information. Among them, the noise correction factor is used to compensate for the errors caused by sensor noise and environmental interference (such as workshop dust and light fluctuations), and it can be dynamically estimated through Gaussian filtering, wavelet transform, or Kalman filtering.

[0071] S3. Construct a defect feature matrix according to the multispectral feature values and the noise correction factor, perform path planning according to the defect feature matrix, and generate a control instruction according to the planned path.

[0072] Preferably, for the aluminum material to be sprayed, the surface image of the aluminum material to be sprayed can be obtained first through a high-resolution camera, and a two-dimensional Cartesian coordinate system can be established according to the surface image of the aluminum material to be sprayed. The defect feature matrix is constructed based on the established two-dimensional Cartesian coordinate system.

[0073] Specifically, by constructing a defect feature matrix function where λ i represents the importance weight of the i-th spectral channel in defect detection, and the numerical range is usually [0,1]. Different wavelengths (such as visible light, near-infrared, and ultraviolet) have different sensitivities to specific defects (such as scratches, oxidation spots, and coating unevenness), and it can be determined by experiments or deep learning optimization. For example, it can be initialized through transfer learning (such as a ResNet pre-trained model), and then fine-tuned through defect samples; I i (x,y) represents the intensity value of the reflected or transmitted light of the i-th spectral channel at the coordinate (x,y), and a high-precision industrial camera (such as a 3D line scanner) or a multispectral sensor array can be used to capture multimodal information such as the surface texture, depth, and spectral reflectance of the aluminum material; ε represents the noise correction factor, n represents the total number of multispectral channels, and 4-8 channels are commonly used in industrial detection, covering the visible light to near-infrared band. In this embodiment, it includes channels such as 530nm (green light), 850nm (near-infrared), and 1450nm (short-wave infrared).

[0074] For each pixel coordinate (x, y), the defect feature matrix can be weighted and fused according to the defect feature matrix function to generate a comprehensive defect feature map.

[0075] Here, the defect feature matrix is obtained through multi-spectral eigenvalue fusion and noise suppression (through a noise correction factor), achieving a robust expression of defect features, which can provide a high signal-to-noise ratio input for subsequent detection and recognition of aluminum surface defects based on neural networks.

[0076] Environmental noise has a certain impact on the recognition of aluminum surface defects. If environmental noise is ignored during the recognition of aluminum surface defects, it will not only reduce the accuracy of aluminum surface defect recognition but also degrade the subsequent aluminum surface spraying quality. Therefore, it is crucial to correct the aluminum surface feature information according to environmental noise before planning the robot spraying path and setting corresponding spraying parameters.

[0077] In this embodiment, by fusing multi-spectral eigenvalues and compensating and correcting the multi-spectral eigenvalues through a noise correction factor, it is beneficial to improve the accuracy of aluminum surface defect recognition. On this basis, path planning is performed according to the constructed defect feature matrix, and control instructions are generated based on the planned path. The robot control method takes into account error factors brought by environmental noise such as sensor noise and environmental interference (such as workshop dust and light fluctuations), which can improve the accuracy of aluminum surface spraying parameters and path planning, making it more in line with the actual situation and improving the spraying quality.

[0078] S4. Control the robot to act according to the control instructions to perform spraying treatment on the aluminum surface.

[0079] Specifically, the robot responds to the control instructions and performs corresponding actions according to the spraying parameters in the control instructions to achieve spraying treatment on the aluminum surface. The spraying parameters in the control instructions include but are not limited to the nozzle movement speed, nozzle angle, and spraying pressure.

[0080] In summary, the robot control method obtains a noise correction factor, constructs a defect feature matrix according to multi-spectral eigenvalues and the noise correction factor, and performs path planning according to the defect feature matrix. It takes into account the influence of different-dimensional noises on the accuracy of aluminum surface defect recognition and can improve the accuracy of aluminum surface defect recognition. On this basis, the aluminum surface defect information is recognized based on the defect feature matrix, and then path planning is performed, which can better perform spraying treatment on the aluminum surface, thereby improving the spraying quality.

[0081] In one of the embodiments, in step S2, as Figure 2 shown, the specific method for obtaining the noise correction factor according to multi-source noise interference information includes:

[0082] S21. Obtain the noise correction factor at the target position at time t-1, and obtain the time-domain recursion term according to the noise correction factor. Specifically, the time-domain recursion term improves the dynamic response ability of the noise correction factor by utilizing the time-domain information.

[0083] S22. Perform multi-scale noise decomposition on the defect feature matrix at the target position, and obtain the multi-scale decomposition term according to the decomposition result.

[0084] Performing multi-scale noise decomposition on the defect feature matrix at the target position and using the multi-scale noise separation technology can distinguish high-frequency random noise from low-frequency defect features.

[0085] S23. Obtain the defect confidence at the target position, and obtain the defect confidence weight term according to the defect confidence.

[0086] The defect confidence can be the defect probability output by a constructed neural network model for identifying aluminum surface defects, such as the YOLOv7-Transformer model, etc. Here, the defect confidence is used as one of the parameters of the final noise correction factor, integrating the defect detection result feedback mechanism, and can dynamically adjust the noise suppression intensity.

[0087] S24. Obtain the environmental disturbance vector at time t, and obtain the environmental disturbance compensation term according to the environmental disturbance vector.

[0088] The environmental disturbance vector includes sensor data such as temperature, humidity, light intensity, and mechanical vibration collected in real time. Through multi-physical field coupling compensation, it can improve the adaptability of the industrial site.

[0089] S25. Obtain the noise correction factor at time t according to the time-domain recursion term, the multi-scale decomposition term, the confidence weight term, and the environmental disturbance compensation term.

[0090] Here, the noise correction factor at time t is obtained through the time-domain recursion term, the multi-scale decomposition term, the confidence weight term, and the environmental disturbance compensation term. The finally obtained noise correction factor integrates multi-dimensional factors such as defect detection feedback, environmental perception, and multi-scale analysis, and can realize the spatio-temporal adaptive optimization of the noise correction term. Compared with the traditional Gaussian filtering method, it can improve the accuracy of aluminum surface defect detection and recognition.

[0091] Preferably, the noise correction factor ε(t,x,y) at time t = TD + MRA·CW + EDC; where TD = ρ·ε t-1 represents the time-domain recursion term, represents the scale decomposition term, represents the confidence weight term, represents the environmental disturbance compensation term, ρ represents the time decay factor, ε t-1denotes the noise correction factor of the target position at time t-1, ω k denotes the weight coefficient of the k-th scale decomposition, D k (x,y) denotes the defect feature matrix of the target position (x,y), Ψ k denotes the decomposition function of the k-th scale, Ψ k (D k (x,y)) denotes the decomposition result of the defect feature matrix of the target position at the k-th scale, K denotes the number of frequency bands of the scale decomposition, which is used to improve the resolution ability for complex noise, C d denotes the defect confidence at the target position, e denotes the natural constant, T denotes the defect confidence activation threshold, η denotes the defect confidence sensitivity coefficient, γ denotes the environmental perturbation compensation coefficient, E env (t) denotes the environmental perturbation vector, σ denotes the normalization parameter.

[0092] Specifically, the time-domain recursion term introduces the recursive filtering idea to suppress instantaneous interference, and the time decay factor is used to control the retention ratio of the historical noise estimation, or to control the influence of the noise correction factor at the previous moment on the noise correction factor at the current moment, usually between 0.8 and 0.95. ω k is used to balance the importance of different scale decompositions and capture the noise characteristics at different scales. The defect confidence is used to improve the inspection accuracy.

[0093] The defect confidence sensitivity coefficient is used to control the influence of the defect confidence on the noise correction factor and enhance the adaptability to defects with different confidences; the defect confidence activation threshold is used to distinguish high-confidence and low-confidence defects, optimize the recall rate and precision of defect detection, and its value is usually between 0.6 and 0.8. When the defect confidence is greater than the activation threshold, the noise correction intensity is reduced to avoid loss of defect features; the environmental perturbation compensation coefficient is used to control the influence of the environmental perturbation compensation term on the noise correction factor and improve the robustness of the algorithm in different environments; the environmental perturbation vector reflects the perturbation level of the current environmental state and is used to dynamically adjust the algorithm parameters to adapt to environmental changes.

[0094] For the normalization parameter, it can be set according to the formula σ = 0.1·max(E env ), which can solve the limitations of traditional linear compensation through non-linear environmental mapping. For multi-scale noise decomposition, the original signal can be decomposed into K = 6 frequency bands by the dual-tree complex wavelet transform and the energy entropy of each layer can be calculated.

[0095] For the defect confidence sensitivity coefficient, preferably, an incremental learning mechanism is established to update the parameters every 5 seconds, 15 seconds, 30 seconds or one minute. Among them, Loss function The defect confidence sensitivity coefficient is usually between 0.1 and 0.3, and its update process is achieved through backpropagation of gradients. When the confidence of the defect area fluctuates greatly, the defect confidence sensitivity coefficient automatically increases to quickly respond to feature changes.

[0096] α1 represents the learning rate coefficient, and its range can be set to 0.001 - 0.01, which determines the parameter adjustment amplitude for each iteration. D1 and D2 respectively represent the defect feature matrix after noise correction and the noiseless defect matrix calibrated in the laboratory. The dimension of the defect feature matrix after noise correction is the same as the original data. The noiseless defect matrix calibrated in the laboratory is used as a supervision signal to guide the denoising direction and can be periodically obtained through high-precision offline detection equipment. λ” represents the regularization coefficient, which is a weight factor used to balance the reconstruction accuracy and feature retention (recommended 0.05 - 0.2). If its value is too large, it will cause the defect edge to be blurred; if it is too small, it will not be able to suppress artifacts. represents the defect confidence gradient, which can be understood as the spatial gradient of the defect probability. ||·||1 represents the L1 norm, which is used to constrain the sparsity of the defect gradient field, suppress over-smoothing in the uniform area, and at the same time retain the abrupt change characteristics of the defect boundary. represents the L2 norm.

[0097] By introducing gradient constraints, the loss function can prevent over-smoothing. The defect confidence sensitivity coefficient is dynamically updated based on the incremental learning mechanism, which can further improve the adaptability and robustness of the algorithm.

[0098] The environmental perturbation vector can be obtained by correlating and coupling the environmental temperature and humidity and mechanical vibration. Specifically, E env (t) = β1·ΔT + β2·ΔH + β3·sin(2πft + φ). Where, ΔT, ΔH, and f respectively represent the temperature deviation (the difference between the current temperature and the calibrated reference temperature), the humidity deviation (the difference between the current humidity and the calibrated reference humidity), and the main frequency of mechanical vibration (the main vibration frequency during the operation of the production equipment), φ represents the phase offset term, which can be understood as the phase difference of the vibration signal relative to the sampling time reference, and β1, β2, and β3 respectively represent the influence weight coefficients of the temperature deviation, humidity deviation, and the main frequency of mechanical vibration, which can be set by technicians according to experience or estimated online based on ridge regression. For example, they can be set as β1 = 0.8, β2 = 1.2, and β3 = 0.5 respectively.

[0099] In summary, the noise correction factor algorithm at time t realizes the efficient correction of noise in the defect feature matrix on the aluminum surface by introducing a four-dimensional collaborative mechanism of time-domain recursion, multi-scale decomposition, defect confidence weight, and environmental perturbation compensation. At the same time, the environmental perturbation compensation term enhances the adaptability of the algorithm in different environments and improves the overall performance of defect detection.

[0100] In one of the embodiments, as Figure 3 shown, in step S3, the specific method for path planning according to the defect feature matrix includes:

[0101] S31, extracting the spatial features of the defect feature matrix through CNN and extracting the temporal features of the defect feature matrix through LSTM.

[0102] Specifically, an improved YOLOv7+Transformer model is constructed. In the field of computer vision, the YOLO (You Only Look Once) series of models are widely popular due to their fast and accurate characteristics. YOLOv7 introduces a variety of improvements, including using the Transformer architecture to enhance the performance of object detection. At the same time, combining CNN (Convolutional Neural Network) and LSTM (Long Short-Term Memory Network) into the Transformer model can further improve the complexity and performance of the model, especially when dealing with video sequences and temporal data.

[0103] Among them, CNN is used to extract spatial features, and LSTM is used to extract temporal features. Combining the output of LSTM with the feature map of CNN and inputting it into the Transformer module, a special attention mechanism can be designed to consider both spatial and temporal information simultaneously.

[0104] The spatial features of the extracted defect feature matrix can be expressed as CNN(D(x,y)). Mainly, through the improved YOLOv7 architecture, the detection of micro-defects is enhanced through residual connections, and the spatial features of high-level semantic feature tensors such as texture, color, and shape are output.

[0105] The temporal features of the extracted defect feature matrix can be expressed as LSTM(▽D). The long short-term memory network is used to perform temporal modeling on the gradient sequence. Its input is the spatio-temporal evolution sequence of the defect edge (such as the defect diffusion path during the spraying process), and the output is the hidden state vector of the dynamic change trend of the defect, which is used to predict the potential defect expansion area. represents the spatial gradient of the defect feature matrix, which characterizes the mutation features of the surface defect edge and is used to capture the contour information of defects such as cracks and pits.

[0106] S32, fusing the spatial features and the temporal features to obtain the defect probability at the target position. Specifically, the defect probability It represents a fusion operation, which is used to weightedly fuse the output features of the CNN and the LSTM. The Sigmoid activation function is used to map the fused features to the interval [0, 1], and the output value represents the probability of a defect existing at the target position (x, y). When the defect probability is greater than the preset probability threshold, it is determined that the target position is a defect area, so as to make corresponding adjustments to the spraying parameters and / or dynamically correct the spraying path.

[0107] Specifically, parameter correction values corresponding to different defect information (including but not limited to defect types and sizes) can be preset first, and then according to the defect information corresponding to the determined defect area, the corresponding parameter correction values are called to dynamically correct the spraying parameters in real time.

[0108] S33, obtain the trajectory tracking error term according to the actual path and the reference path, obtain the defect gradient field according to the defect probability, and obtain the defect avoidance term according to the defect gradient field.

[0109] S34, obtain the path optimization objective function according to the trajectory tracking error term and the defect avoidance term, and perform path planning according to the path optimization objective function.

[0110] Preferably, the path optimization objective function

[0111] Among them, respectively represent the trajectory tracking error term and the defect avoidance term, P ref 、P act respectively represent the reference path and the actual path, ||·|| represents the Euclidean distance norm, α represents the trajectory tracking weight coefficient, which is used to adjust the priority of the tracking accuracy in the overall objective, M d represents the defect probability, represents the defect gradient field of the defect probability in the x direction, β represents the defect avoidance weight coefficient, t0 and t1 respectively represent the start and end times of the spraying operation, and C represents the integral path, usually the motion profile planned for the end effector of the robot.

[0112] Specifically, the reference path is generated by process specifications or initial path planning, and includes the ideal position (such as (x, y, z) coordinates) and motion parameters (speed, angle, etc.) of the end effector of the spraying robot. The larger the trajectory tracking weight coefficient, the more the system tends to strictly track the reference trajectory, but it may sacrifice the ability to avoid defects. For the defect gradient field, at high gradient values, it can prompt the robot to adjust the path to avoid defect-dense areas (such as edge burrs or oxidation patches). The larger the defect avoidance weight coefficient, the more the path planning tends to bypass defects, but it may increase the trajectory tracking error.

[0113] It represents the line integral of the defect gradient along the integral path, quantifying the cumulative effect of the defect distribution around the path. If the integral value is large, it indicates that the path needs to shift towards the low-gradient region.

[0114] The path optimization objective function can be understood as a multi-objective reward function. In this embodiment, a defect gradient sensitive term is introduced for the industrial scenario. By directly embedding the defect detection result into the path optimization, it is more adaptable to the complex surface defect distribution than traditional obstacle avoidance algorithms (such as the artificial potential field method).

[0115] In this embodiment, by fusing the real-time defect gradient and the trajectory tracking error, the collaborative optimization of the spraying quality and process stability is achieved.

[0116] In one of the embodiments, as Figure 4 shown, in step S3, the specific method for generating control instructions according to the planned path includes:

[0117] S35, construct an RBFNN model, train the RBFNN model, and obtain the weight matrix.

[0118] The weight matrix is the trainable parameter of the RBFNN (Radial Basis Function Neural Network) model, used to map the input variables to the spraying parameter space. Specifically, transfer learning can be used to initialize the weight matrix, and the cloud ResNet50 pre-trained model is used to extract defect features.

[0119] The output layer weight of the RBFNN is updated in real time through supervised learning (such as the gradient descent method or the pseudo-inverse method), so that the spraying parameters can adapt to the defect distribution, path error, and environmental disturbance. For example: when high-density defects are detected, the RBFNN automatically increases the local spraying thickness; when the path optimization objective function shows a trajectory deviation, the moving speed of the spray gun is dynamically adjusted to compensate for the error.

[0120] S36, use the environmental disturbance vector, defect probability, and path optimization objective function as the input of the RBFNN model for high-dimensional space mapping;

[0121] S37, construct a PID controller, and obtain the PID control error according to the PID controller.

[0122] Specifically, the PID control error can be understood as the real-time error between the actual coating thickness and the target thickness. The theoretical thickness can be predicted through the digital twin model, and then compared with the measured value to generate the PID control error. The PID control error is transmitted to the PID controller to dynamically adjust the spraying speed and pressure.

[0123] S38, obtain the comprehensive spraying parameter vector according to the RBFNN model and the PID control error, and generate control instructions according to the comprehensive spraying parameter vector.

[0124] Preferably, the comprehensive spraying vector is obtained according to the formula f = RBFNN(W T ·[M d , J path , E env ) + PID(Δe); where RBFNN(W T ·[M d , J path , E env ) and PID(Δe) respectively represent the RBFNN model and the PID control error, W represents the weight matrix, and W T represents the transpose of the weight matrix.

[0125] During the actual spraying process on the aluminum surface, when a scratch defect is detected in a certain area, the path deviation = 2MM, and the humidity = 70%, the RBFNN calculates the optimal spraying parameters according to the input vector: the spraying speed is 0.4 m / s, and the spraying pressure is 0.3 MPa. At the same time, the PID controller fine-tunes the height of the spray gun to compensate for the trajectory deviation, and finally realizes thickened spraying in the defect area to avoid paint film defects caused by environmental factors.

[0126] Combined with the PID controller, the RBFNN realizes the multi-objective balance of spraying uniformity, defect coverage rate and operation efficiency. That is to say, the comprehensive spraying vector can be understood as a set of optimized spraying parameters, including but not limited to spraying speed, angle, pressure and spraying radius. For example, in the convex / depressed area on the aluminum surface, the combined spraying vector triggers variable-angle spraying, and reduces the non-uniformity of thickness through arc radius optimization. Specifically, when the convex height h = 3mm, the spraying speed is triggered to decrease from v = 0.5m / s to v = 0.3m / s, which can improve the coating thickness uniformity; when the humidity rises from 40% to 70%, the spraying pressure is automatically adjusted from 3 bar to 2.5 bar, which can reduce the sagging defect.

[0127] In this embodiment, by combining the non-linear compensation of the RBFNN and the fast response of the PID, not only the multi-objective optimization of spraying quality, efficiency and cost is realized, but also the adaptive adjustment of spraying parameters can be realized, reducing the non-uniformity of spraying on the aluminum surface, thereby improving the spraying quality.

[0128] As Figure 5 shown, an embodiment of the present invention also provides a robot control system for aluminum surface spraying treatment, which is used to implement the described robot control method, and includes a multi-spectral module, a correction factor acquisition module, a path planning module and a robot control module.

[0129] The multispectral module is used to obtain the multispectral information of the aluminum surface and extract the multispectral feature values according to the multispectral information; the correction factor acquisition module is used to obtain the multi-source noise interference information and obtain the noise correction factor according to the multi-source noise interference information.

[0130] The multispectral module includes but is not limited to multispectral sensors and high-precision industrial cameras (such as 3D line array scanning instruments), and the multispectral information includes multi-modal information such as the texture, depth, and spectral reflectivity of the aluminum surface.

[0131] Preferably, as Figure 6 shown, the correction factor acquisition module includes a time-domain recursive term acquisition unit, a multi-scale decomposition unit, a defect confidence acquisition unit, an environmental disturbance acquisition unit, and a correction factor acquisition unit.

[0132] The time-domain recursive term acquisition unit is used to obtain the noise correction factor at the target position at time t-1 and obtain the time-domain recursive term according to the noise correction factor; the multi-scale decomposition unit is used to perform multi-scale noise decomposition on the defect feature matrix at the target position and obtain the multi-scale decomposition term according to the decomposition result; the defect confidence acquisition unit is used to obtain the defect confidence at the target position and obtain the defect confidence weight term according to the defect confidence.

[0133] The environmental disturbance acquisition unit is used to obtain the environmental disturbance vector at time t and obtain the environmental disturbance compensation term according to the environmental disturbance vector; the correction factor acquisition unit is used to obtain the noise correction factor at time t according to the time-domain recursive term, the multi-scale decomposition term, the confidence weight term, and the environmental disturbance compensation term.

[0134] Preferably, the noise correction factor ε(t,x,y) at time t = TD + MRA·CW + EDC,; where TD = ρ·ε t-1 represents the time-domain recursive term, represents the scale decomposition term, represents the confidence weight term, represents the environmental disturbance compensation term, ρ represents the time decay factor, ε t-1 represents the noise correction factor at the target position at time t-1, ω k represents the weight coefficient of the k-th scale decomposition, D k (x,y) represents the defect feature matrix at the target position (x,y), Ψ k represents the decomposition function of the k-th scale, Ψ k (D k (x,y)) represents the decomposition result of the defect feature matrix at the target position at the k-th scale, K represents the number of frequency bands of the scale decomposition, which is used to improve the resolution ability of complex noise, C dIndicates the defect confidence level at the target position, e represents the natural constant, T represents the defect confidence level activation threshold, η represents the defect confidence level sensitivity coefficient, γ represents the environmental perturbation compensation coefficient, and E env (t) represents the environmental perturbation vector, and σ represents the normalization parameter.

[0135] Here, the noise correction factor algorithm at time t realizes the efficient correction of noise in the defect feature matrix on the aluminum surface by introducing a four-dimensional collaborative mechanism of time-domain recursion, multi-scale decomposition, defect confidence level weight, and environmental perturbation compensation. At the same time, the environmental perturbation compensation term enhances the adaptability of the algorithm in different environments and improves the overall performance of defect detection.

[0136] The path planning module is used to construct a defect feature matrix based on the multi-spectral eigenvalue and the noise correction factor, perform path planning according to the defect feature matrix, and generate a control command according to the planned path; the robot control module controls the robot's actions according to the control command to spray the aluminum surface.

[0137] Preferably, as Figure 7 shown, the path planning module includes a feature extraction unit, a feature fusion unit, an objective function acquisition unit, and a path planning unit.

[0138] The feature extraction unit is used to extract the spatial feature of the defect feature matrix through CNN and the temporal feature of the defect feature matrix through LSTM; the feature fusion unit is used to fuse the spatial feature and the temporal feature to obtain the defect probability at the target position.

[0139] The objective function acquisition unit is used to obtain the trajectory tracking error term according to the actual path and the reference path, obtain the defect gradient field according to the defect probability, and obtain the defect avoidance term according to the defect gradient field; the path planning unit is used to obtain the path optimization objective function according to the trajectory tracking error term and the defect avoidance term, and perform path planning according to the path optimization objective function.

[0140] Specifically, the defect probability It represents a fusion operation, which is used to weight and fuse the output features of CNN and LSTM. The Sigmoid activation function is used to map the fused features to the interval [0, 1], and the output value represents the probability of a defect existing at the target position (x, y). When the defect probability is greater than the preset probability threshold, it is determined that the target position is a defect area, so as to make corresponding adjustments to the spraying parameters and / or dynamically correct the spraying path. The spatial features of the extracted defect feature matrix can be represented as CNN(D(x, y)). The temporal features of the extracted defect feature matrix can be represented as LSTM(▽D), where ▽D represents the spatial gradient of the defect feature matrix, which characterizes the mutation features of the surface defect edge and is used to capture the contour information of defects such as cracks and pits.

[0141] Path optimization objective function Wherein, respectively represent the trajectory tracking error term and the defect avoidance term, P ref 、P act respectively represent the reference path and the actual path, ||·|| represents the Euclidean distance norm, α represents the trajectory tracking weight coefficient, which is used to adjust the priority of the tracking accuracy in the overall objective, M d represents the defect probability, represents the defect gradient field of the defect probability in the x direction, β represents the defect avoidance weight coefficient, t0 and t1 respectively represent the start and end times of the spraying operation, and C represents the integration path, which is usually the motion profile planned by the end effector of the robot.

[0142] Specifically, the YOLOv7-Transformer model includes an input layer, a decision layer, and an output layer. The input layer is used to collect data in real time, including the defect feature matrix, the spraying trajectory error, and environmental parameters, etc.; the decision layer adopts a hybrid structure of RBF neural network and PID, and obtains the comprehensive spraying vector according to the formula f = RBFNN(W T ·[M d ,J path ,E env ) + PID(Δe); the output layer is used to output the dynamic spraying vector, including but not limited to the spraying speed, angle, pressure, and spray gun height.

[0143] Here, a defect gradient sensitive term is introduced for the industrial scenario. By directly embedding the defect detection result into the path optimization, it is more adaptable to the complex surface defect distribution than traditional obstacle avoidance algorithms (such as the artificial potential field method). The path optimization objective function realizes the collaborative optimization of spraying quality and process stability by fusing the real-time defect gradient and the trajectory tracking error.

[0144] In one embodiment, such as Figure 8As shown, the robot control system for the surface spraying treatment of aluminum profiles further includes a spraying weight acquisition module.

[0145] The spraying weight acquisition module obtains the spraying weight according to where k' represents the defect severity gain, which is used to control the influence slope of the defect severity on the spraying weight. When the defect severity gain increases, the response of the weight function to the defect probability becomes steeper, and the high-severity defect area is preferentially strengthened. τ represents the defect activation threshold, which can be understood as the lowest standard for defect recognition. It determines the triggering condition for spraying enhancement. When the defect probability is greater than the defect activation threshold, the weight function increases significantly, and spraying compensation is started. When the defect probability is less than or equal to the defect activation threshold, the weight approaches 0 to avoid redundant spraying. P robot represents the real-time coordinates of the spraying robot in the spraying plane currently, and R spray represents the spraying radius, which is determined by the spray gun model and process parameters. When the spraying point exceeds the spraying radius, the weight approaches 0 and the spraying is invalid. Inside the spraying radius, the weight decreases linearly with the distance from the current position of the spraying robot.

[0146] can be understood as the defect activation factor, and can be understood as the distance attenuation factor between the spraying point (x, y) and the current position of the robot.

[0147] According to the spraying weight is obtained. When the defect probability is greater than the defect activation threshold, the defect severity gain can ensure the thickening of the coating in the defect area by adjusting the spraying compensation intensity; through the online learning of the defect severity gain and the defect activation threshold (such as LSTM update), it can adapt to different surface defect characteristics of aluminum profiles.

[0148] Here, the spraying parameters of the system can be dynamically coupled and optimized through the spraying weight. For example, the spraying weight is proportional to the spraying pressure. The higher the weight in the defect area, the greater the increase in pressure, which is used to achieve enhanced defect coverage and atomization uniformity; the spraying weight is inversely proportional to the spraying speed. The higher the spraying weight, the slower the speed, ensuring an extended residence time in the defect area to achieve uniform coating thickness and sag control; for the spraying angle, the spray gun can be deflected towards the defect direction through the weight gradient to cover the spraying edge and reduce splashing; for the spraying distance, the distance between the spray gun and the support in the defect area can be shortened to enhance the deposition rate, control the adhesion and orange peel texture.

[0149] The spraying weight can be used as the feature of the input layer of the neural network to realize the non-linear mapping of the spraying pressure and speed; at the same time, the spraying weight can be incorporated into the path optimization objective function to achieve the Pareto optimality of defect avoidance and efficiency.

[0150] In one embodiment, as Figure 9 shown, the robot control system for aluminum surface spraying treatment further includes a multi-objective optimization module. The multi-objective optimization module is used to pass through a multi-objective optimization function

[0151] wherein, H real_i , H target_i respectively represent the actual coating thickness and the target coating thickness at the i-th sampling point, N represents the number of preset sampling points, represents the coating uniformity term. The actual coating thickness can be collected in real time by a laser thickness gauge or a vision sensor, and the target coating thickness can be dynamically adjusted according to parameters such as workpiece material properties, environmental temperature and humidity.

[0152] D' represents the defect set, and W' j represents the spraying weight of the j-th defect area. Specifically, represents the real-time coordinates of the spraying robot in the j-th defect area, represents the defect probability of the j-th defect area, (x j , y j ) represents the coordinate position of the j-th defect area.

[0153] lg(1 + W' j ) represents the non-linear strengthening factor, and its logarithmic function converts the spraying weight into a non-linear growth term, enhancing the coverage priority of high-weight areas and ensuring that the coverage of defect areas is appropriately emphasized. represents the defect coverage strengthening term of the weight coefficient, which is used to adjust the importance of defect coverage in the whole optimization, generally between 0.1 and 0.3.

[0154] For the non-linear strengthening factor, when W' j is small, lg(1 + W' j ) grows smoothly, avoiding over-strengthening of low-priority areas; when W' j is large, the growth rate of the logarithmic function slows down, preventing excessive spraying caused by too high a weight. That is to say, the non-linear strengthening factor balances defect coverage and spraying efficiency through non-linear conversion, ensuring that defects are preferentially repaired while meeting coating uniformity.

[0155] T total represents the operation efficiency term, which can be understood as the total time of the whole operation process, determined by the spraying path length and the robot moving speed. represents the weight coefficient of the operation efficiency term, which is used to adjust the importance of the operation efficiency in the whole optimization. A smaller value of the operation efficiency term means that the operation is completed faster and more efficiently. The weight coefficient of the operation efficiency term is generally between 0.05 and 0.1.

[0156] Θ = [v, p, Δt] ∈ Ω feasible , which can be understood as a parameter vector containing three parameters: spraying speed v, spraying pressure p, and time interval Δt. Ω feasible represents the feasible region defined by the robot dynamics to ensure the normal operation of the robot and avoid dynamically infeasible situations (such as the robot becoming unstable due to excessive speed). For the spraying speed, spraying pressure, and time interval, they can be set according to the actual situation. In one aluminum spraying scenario, the spraying speed range can be set to 0.1 - 1.2 m / s, and the spraying pressure range can be set to 0.3 - 0.8 MPa.

[0157] It should be noted that the path optimization objective function focuses on the dynamic adjustment of the spraying path, and the multi-objective optimization function focuses on the system-level parameter decision-making. It is necessary to find the Pareto optimal solution set among multiple conflicting objectives. In actual control, the weight coefficients of the trajectory tracking error term and the defect avoidance term in the path optimization objective function can be dynamically adjusted by the multi-objective optimization results to form a two-layer optimization architecture. The outputs of the multi-objective optimization function, such as spraying speed, pressure, etc., can be used as constraint conditions and input into the path optimization objective function to affect its feasible solution space. The real-time data of the multi-objective optimization function (such as the actual trajectory error) can be fed back to the multi-objective function to update the objective weights or constraint boundaries. Preferably, non-dominated solution sets can be generated through algorithms such as NSGA-II, and a compromise solution that satisfies the constraints of the path optimization objective function can be selected from them. Or the hard constraints of the path optimization objective function can be transformed into penalty terms of the multi-objective function to allow limited deviation in exchange for the global optimum to coordinate the path optimization objective function and the multi-objective optimization function.

[0158] The multi-objective optimization function simultaneously considers uniformity, defect coverage, and operation efficiency. Combining with the robot dynamics constraints, it can find the best balance solution among multiple conflicting objectives, realizing the collaborative optimization of spraying quality, efficiency, and cost. Its innovation lies in the dynamic coupling of defect detection data and spraying parameters, while taking into account the real-time and robustness requirements of industrial scenarios.

[0159] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in this specification.

[0160] The above-described embodiments merely represent several implementation manners of the present invention. The description thereof is relatively specific and detailed, but it should not be construed as a limitation on the scope of the invention patent. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all fall within the protection scope of the present invention. Therefore, the protection scope of the present invention patent shall be subject to the appended claims.

Claims

1. A robot control method for aluminum surface spraying, characterized in that: The steps include: Acquire multispectral information of the aluminum surface, and extract multispectral feature values ​​according to the multispectral information; Acquire multi-source noise interference information, and acquire a noise correction factor according to the multi-source noise interference information; Construct a defect feature matrix based on multi-spectral eigenvalues ​​and noise correction factors, perform path planning based on the defect feature matrix, and generate control instructions based on the planned path; The robot is controlled according to the control instructions to spray the surface of the aluminum material.

2. A robot control method for aluminum surface spraying as claimed in claim 1, characterized in that: The specific method of obtaining the noise correction factor according to the multi-source noise interference information includes: Obtain the noise correction factor of the target position at time t-1, and obtain the time domain recursive term according to the noise correction factor; Perform multi-scale noise decomposition on the defect feature matrix of the target position, and obtain multi-scale decomposition items according to the decomposition results; Obtain the defect confidence of the target position, and obtain the defect confidence weight item according to the defect confidence; Obtaining an environmental disturbance vector at time t, and obtaining an environmental disturbance compensation item according to the environmental disturbance vector; The noise correction factor at time t is obtained according to the time domain recursive term, the multi-scale decomposition term, the confidence weight term and the environmental disturbance compensation term.

3. A robot control method for aluminum surface spraying as claimed in claim 2, characterized in that: The specific methods for path planning based on the defect feature matrix include: The spatial features of the defect feature matrix are extracted through CNN, and the temporal features of the defect feature matrix are extracted through LSTM; The spatial features and temporal features are integrated to obtain the defect probability at the target location; Obtain a trajectory tracking error term according to the actual path and the reference path, obtain a defect gradient field according to the defect probability, and obtain a defect avoidance term according to the defect gradient field; The path optimization objective function is obtained according to the trajectory tracking error term and the defect avoidance term, and the path planning is performed according to the path optimization objective function.

4. A robot control method for aluminum surface spraying as claimed in claim 3, characterized in that: The specific method of generating control instructions according to the planned path includes: Build the RBFNN model, train the RBFNN model, and obtain the weight matrix; The environmental disturbance vector, defect probability and path optimization objective function are used as the input of the RBFNN model for high-dimensional space mapping; Construct a PID controller and obtain the PID control error based on the PID controller; The comprehensive spraying parameter vector is obtained according to the RBFNN model and the PID control error, and the control instruction is generated according to the comprehensive spraying parameter vector.

5. A robot control method for aluminum surface spraying as claimed in claim 4, characterized in that: Noise correction factor ε(t,x,y) at time t = TD + MRA·CW + EDC; Where TD = ρ·ε t-1 represents the time domain recursive term, represents the scale decomposition term, represents the confidence weight term, represents the environmental disturbance compensation term, ρ represents the time attenuation factor, ε t-1 Indicates that time t-1 is the noise correction factor of the target position, ω k represents the weight coefficient of the k-th scale decomposition, D k (x,y) represents the defect feature matrix of the target position (x,y), Ψ k (D k (x,y)) represents the decomposition result of the defect feature matrix of the target position under the kth scale. K represents the number of frequency bands of scale decomposition, C d represents the defect confidence at the target position, e represents the natural constant, T represents the defect confidence activation threshold, η represents the defect confidence sensitivity coefficient, γ represents the environmental disturbance compensation coefficient, and E env (t) represents the environmental disturbance vector, and σ represents the normalization parameter.

6. A robot control method for aluminum surface spraying as claimed in claim 5, characterized in that: Path optimization objective function in, denote the trajectory tracking error term and the defect avoidance term respectively, P ref , P act denote the reference path and the actual path respectively, ||·|| denotes the Euclidean distance norm, α denotes the trajectory tracking weight coefficient, and M d represents the defect probability, represents the defect gradient field of the defect probability in the x direction, β represents the defect avoidance weight coefficient, t0 and t1 represent the start and end time of the spraying operation respectively, and C represents the integral path.

7. A robot control method for aluminum surface spraying as claimed in claim 6, characterized in that: According to the formula f = RBFNN (W T ·[M d ,J path ,E env ])+PID(Δe) to obtain the comprehensive spraying vector; Among them, RBFNN(W T ·[M d ,J path ,E env ]) and PID(Δe) represent the RBFNN model and PID control error respectively, W represents the weight matrix, W T represents the transpose of the weight matrix.

8. A robot control system for aluminum surface spraying, used to implement the robot control method according to any one of claims 1 to 7, characterized in that: include: The multispectral module is used to obtain the multispectral information of the aluminum surface and extract the multispectral characteristic value according to the multispectral information; A correction factor acquisition module is used to obtain multi-source noise interference information and obtain a noise correction factor according to the multi-source noise interference information; A path planning module is used to construct a defect feature matrix based on multi-spectral feature values ​​and noise correction factors, perform path planning based on the defect feature matrix, and generate control instructions based on the planned path; The robot control module controls the robot's movements according to control instructions and sprays the aluminum surface.

9. A robot control system for aluminum surface spraying as claimed in claim 8, characterized in that: The correction factor acquisition module includes: A time domain recursive item acquisition unit, used to acquire a noise correction factor of a target position at time t-1, and acquire a time domain recursive item according to the noise correction factor; A multi-scale decomposition unit is used to perform multi-scale noise decomposition on the defect feature matrix at the target position and obtain multi-scale decomposition items according to the decomposition results; A defect confidence acquisition unit, used to acquire the defect confidence of the target position, and acquire a defect confidence weight item according to the defect confidence; An environmental disturbance acquisition unit, used to acquire an environmental disturbance vector at time t, and acquire an environmental disturbance compensation item according to the environmental disturbance vector; The correction factor acquisition unit is used to obtain the noise correction factor at time t according to the time domain recursive term, the multi-scale decomposition term, the confidence weight term and the environmental disturbance compensation term.

10. A robot control system for aluminum surface spraying as claimed in claim 9, characterized in that: The path planning module includes: A feature extraction unit, used to extract the spatial features of the defect feature matrix through CNN and the temporal features of the defect feature matrix through LSTM; A feature fusion unit is used to fuse spatial features and temporal features to obtain the defect probability at the target location; An objective function acquisition unit, used to acquire a trajectory tracking error term according to an actual path and a reference path, acquire a defect gradient field according to a defect probability, and acquire a defect avoidance term according to the defect gradient field; The path planning unit is used to obtain a path optimization objective function according to a trajectory tracking error term and a defect avoidance term, and perform path planning according to the path optimization objective function.

Images