Automobile Parts Spraying Control Method and System

Abstract

The present invention provides a spraying control method and system for automotive parts, which relates to the technical field of automotive manufacturing. The method includes: collecting spraying environment parameters in real time, generating a comprehensive digital model including environmental parameters and part characteristics, using a clustering and partitioning algorithm to intelligently partition the surface of automotive parts into multiple sub-regions; based on each sub-region, combining the current environmental parameters, using an intelligent agent trained by a reinforcement learning algorithm to generate optimal spraying parameters for each sub-region, based on the optimal spraying parameters and the spatial relationship between sub-regions, using an ant colony path planning algorithm to optimize the global spraying sequence and generate an optimal spraying path; a spraying robot sprays to generate a real-time quality evaluation result of the spraying, using a pre-constructed fuzzy logic controller to adjust the spraying parameters to obtain the sprayed parts, integrating the full-cycle comprehensive data set, and iteratively updating the spraying control parameters.

Description

Technical Field

[0001] The present invention relates to the technical field of automobile manufacturing, and particularly to a spraying control method and system for automobile parts. Background Art

[0002] The spraying quality of automobile parts directly affects the appearance and anti-corrosion performance of automobiles, and is a key link in the automobile manufacturing process. Traditional spraying control methods for automobile parts mainly rely on manual experience and fixed parameter settings, which are prone to problems.

[0003] In the prior art, the setting of spraying parameters lacks flexibility, resulting in unstable spraying quality; the manual detection method is highly subjective and the quality control accuracy is insufficient; the trial-and-error method is usually adopted, and the optimization process is inefficient; the shapes of automobile parts are complex and diverse, and it is difficult for traditional methods to cope with complex working conditions; there is a lack of data-driven continuous optimization ability, and it is difficult to balance multiple quality indicators.

[0004] In summary, it is necessary to apply artificial intelligence technology combined with intelligent control technology to the spraying process of automobile parts, and develop a more intelligent and adaptive spraying control method. The present invention can solve the problems in the prior art. Summary of the Invention

[0005] An embodiment of the present invention provides a spraying control method and system for automobile parts, which can solve the problems in the prior art.

[0006] In the first aspect of the embodiment of the present invention,

[0007] A spraying control method for automobile parts is provided, including:

[0008] Real-time collecting spraying environment parameters through a temperature and humidity sensor, a pressure sensor and an air quality sensor; obtaining surface spectral information and three-dimensional structure data of the automobile part by using a multi-spectral camera and a three-dimensional structured light scanner; inputting the spraying environment parameters, the surface spectral information and the three-dimensional structure data into a pre-trained multi-modal fusion neural network to generate a comprehensive digital model including environment parameters and part features, and based on the comprehensive digital model, using a clustering and partitioning algorithm to intelligently partition the surface of the automobile part into multiple sub-regions;

[0009] Based on each sub-region, combined with the current environmental parameters, an intelligent agent trained using a reinforcement learning algorithm generates optimal spraying parameters for each sub-region. Based on the optimal spraying parameters and the spatial relationship between sub-regions, an ant colony path planning algorithm is used to optimize the global spraying sequence and generate an optimal spraying path. The optimal spraying parameters and the optimal spraying path are sent to a spraying robot with adaptive control capabilities. During the spraying process, spraying status data is collected in real-time through a high-speed camera and a coating thickness sensor, and the spraying status data is input into a pre-trained convolutional neural network to generate a real-time quality assessment result of the spraying.

[0010] Using a pre-constructed fuzzy logic controller, according to the deviation between the real-time quality assessment result and the preset quality target, the spraying parameters are adjusted until the current automotive component spraying is completed, obtaining a sprayed completed component. The sprayed completed component is scanned and detected to obtain the final spraying quality data. The final spraying quality data, environmental parameters, component characteristics, and spraying parameters are integrated into a full-cycle comprehensive data set. Based on the full-cycle comprehensive data set, through a gradient boosting decision tree algorithm, variables affecting spraying quality are extracted. Based on the variables affecting spraying quality, the network parameters of the multi-modal fusion neural network and the algorithm parameters of the reinforcement learning algorithm are updated.

[0011] In an alternative embodiment,

[0012] The spraying environmental parameters, surface spectral information, and three-dimensional structure data are input into a pre-trained multi-modal fusion neural network to generate a comprehensive digital model containing environmental parameters and component characteristics. Based on the comprehensive digital model, a clustering partition algorithm is used to intelligently partition the surface of the automotive component into multiple sub-regions, including:

[0013] The spraying environmental parameters are collected, and the spraying environmental parameters include temperature data, humidity data, air pressure data, and airborne particulate matter concentration data. The hyperspectral information on the surface of the automotive component is collected, and the hyperspectral information includes spectral reflectance data of multiple bands. The three-dimensional structure data of the automotive component is collected, and the three-dimensional structure data includes point cloud data.

[0014] A multi-modal fusion neural network is constructed, and the multi-modal fusion neural network includes an environmental parameter branch, a spectral information branch, and a three-dimensional structure branch.

[0015] The spraying environmental parameters are input into the environmental parameter branch, and a one-dimensional convolutional neural network is used to process the temporal environmental data to obtain an environmental feature vector. The hyperspectral information is input into the spectral information branch, and a three-dimensional convolutional neural network is used to extract spatial-spectral features to obtain a spectral feature vector. The three-dimensional structure data is input into the three-dimensional structure branch, and a point cloud convolutional network is used to process the point cloud data to obtain a structural feature vector.

[0016] The multi-head self-attention module is used to fuse the environmental feature vector, the spectral feature vector, and the structural feature vector to obtain a fused feature; the fused feature is input into a multi-layer perceptron, and a feature vector of a fixed dimension is output as a comprehensive digital model;

[0017] Based on the comprehensive digital model, the intelligent partitioning algorithm is used to intelligently partition the surface of automotive parts to determine multiple sub-regions.

[0018] In an alternative embodiment,

[0019] The clustering and partitioning algorithm includes:

[0020] Construct a clustering and partitioning algorithm based on K-means clustering; initialize the cluster centers, and use the feature vectors in the comprehensive digital model as the initial data points;

[0021] Calculate the distance from each data point to each cluster center. When calculating the distance, introduce spatial constraints, add spatial coordinate terms to the distance calculation formula, and assign an adjustable weight coefficient; assign each data point to the cluster center with the closest distance; calculate the average value of all data points in each cluster, determine the new cluster center, and update; calculate the local curvature of each data point, and use the local curvature as the cluster feature dimension;

[0022] Repeat the iteration until the change in the position of the cluster center is less than the preset cluster center change threshold to obtain the optimal partitioning scheme;

[0023] Calculate the silhouette coefficient of the optimal partitioning scheme through a similarity measurement method;

[0024] Within the preset range of the number of partitions, repeat the initialization and iterative calculation process to obtain the silhouette coefficients under different numbers of partitions, and select the partitioning scheme with the largest silhouette coefficient as the final intelligent partitioning result.

[0025] In an alternative embodiment,

[0026] Based on each sub-region, combined with the current environmental parameters, the intelligent agent trained by the reinforcement learning algorithm generates the optimal spraying parameters for each sub-region, including:

[0027] Construct a main network based on the deep Q network. The main network includes an input layer, two hidden layers, and an output layer. Among them, the number of nodes in the input layer is the same as the dimension of the state vector, the number of nodes in the output layer is equal to the size of the action space, the hidden layer uses the ReLU activation function, and the output layer uses the linear activation function;

[0028] Determine the state space based on the geometric features, material properties, and environmental parameters of the sub-regions; determine the action space based on the spraying parameters; calculate the weighted sum of the coating quality indicators in the sub-regions based on the preset coating quality indicators, and construct a reward function;

[0029] Based on the state space, action space, and reward function, construct a simulation environment that simulates the spraying process of the sub-regions on the surface of automotive parts;

[0030] Construct an experience pool for storing the experience data obtained from the interaction between the intelligent agent and the simulation environment, initialize the experience pool, create a target network based on the structure of the main network, and calculate the target Q value;

[0031] Adopt the ε-greedy method, set the exploration strategy, initially set the ε value to 0.9, and linearly decay it to 0.1 as the number of training rounds increases;

[0032] The intelligent agent selects spraying parameters according to the current sub-region state and the exploration strategy, interacts with the simulation environment, obtains the reward and the next sub-region state, and calculates the temporal difference error based on the current sub-region state, the selected spraying parameters, the obtained reward, the next sub-region state, and the target Q value calculated by the target network;

[0033] Take the current sub-region state, the selected spraying parameters, the obtained reward, the next sub-region state, and the temporal difference error as an experience data, store it in and update the experience pool; allocate sampling probabilities for each experience data based on the temporal difference error in the experience pool, select batch data based on the sampling probabilities, and calculate the loss function based on the batch data, the main network, and the target network;

[0034] Minimize the loss function by the gradient descent method, update the parameters of the main network, assign the weighted average of the main network parameters to the corresponding parameters of the target network, update the parameters of the target network, add Gaussian noise to the parameters of the main network and the target network, and continuously iterate until the preset number of training rounds is reached, determine the trained main network, input the current state of each sub-region into the trained main network, and input the optimal spraying parameter combination of the sub-region.

[0035] In an alternative embodiment,

[0036] Based on the optimal spraying parameters and the spatial relationship between the sub-regions, optimize the global spraying order using the ant colony path planning algorithm, and generate the optimal spraying path, including:

[0037] Generate a node corresponding to each sub-region, determine the edges between the nodes based on the spatial relationship between the sub-regions, and determine the graph structure based on the nodes and edges;

[0038] Based on the spatial distance between sub-regions, the optimal spraying parameters corresponding to each sub-region, and the surface curvature change of the sub-region, calculate the edge weight, where the spatial distance is determined based on the Euclidean distance between adjacent sub-regions, the similarity degree between the corresponding vectors of the optimal spraying parameters of two sub-regions is determined by calculating the cosine similarity to obtain the spraying parameter similarity, the curvature between adjacent sub-regions is calculated to determine the curvature difference, and the surface curvature change is obtained;

[0039] Initialize the ant population and the corresponding pheromone matrix according to the preset number of ant populations. Each ant randomly selects a sub-region as the starting sub-region;

[0040] Based on the edge weight, determine the heuristic information. Based on the pheromone matrix, determine the pheromone concentration. According to the heuristic information and the pheromone concentration, calculate the state transition probability to determine the probability of each ant selecting the next sub-region, select the sub-region corresponding to the maximum selection probability, and move the ant to the corresponding sub-region;

[0041] After the ant moves, update the local pheromone of the edge passed during the movement. The local pheromone is updated based on the corresponding pheromone concentration in the pheromone matrix, the preset pheromone evaporation coefficient, and the preset initial pheromone concentration;

[0042] Repeat the steps of sub-region selection and local pheromone update until all ants complete a full path. Based on the spatial distance between sub-regions and the time distance of spray gun adjustment, determine the path length of the ant. Based on the spraying parameter similarity and curvature change between adjacent sub-regions, determine the spraying quality score corresponding to the ant. Update the global pheromone based on the path length and the spraying quality score. Repeat until the preset maximum number of iterations is reached, select the spraying path corresponding to the maximum spraying quality score, and determine the optimal spraying path.

[0043] In an alternative embodiment,

[0044] The path length is calculated by the following formula:

[0045] ;

[0046] where L represents the total path length, m represents the number of sub-regions included in the path, i represents the index of the sub-region in the path, r(·) represents the sub-region corresponding to the index in the path, d(·, ·) represents the spatial distance between sub-regions, T represents the time coefficient required to adjust the spraying parameters, P(·) represents the optimal spraying parameter vector, and ||·|| represents calculating the Euclidean distance;

[0047] The spraying quality score is calculated by the following formula:

[0048] ;

[0049] Among them, Q represents, λ represents the weight coefficient, sim(·) represents calculating the similarity of spraying parameters, c(·) represents calculating the average curvature, and |·| represents the absolute value of the curvature difference.

[0050] In an alternative embodiment,

[0051] Adopt a pre-constructed fuzzy logic controller to adjust the spraying parameters according to the deviation between the real-time quality assessment result and the preset quality target until the spraying of the current automotive part is completed, obtain the sprayed part, scan and detect the sprayed part to obtain the final spraying quality data, integrate the final spraying quality data, environmental parameters, part characteristics and spraying parameters into a full-cycle comprehensive data set, and based on the full-cycle comprehensive data set, extract the spraying quality influencing variables through the gradient boosting decision tree algorithm. Based on the spraying quality influencing variables, update the network parameters of the multi-modal fusion neural network and the algorithm parameters of the reinforcement learning algorithm, including:

[0052] According to the deviation between the real-time quality assessment result and the preset quality target, adjust the spraying parameters through a pre-constructed fuzzy logic controller until the spraying of the current automotive part is completed, including:

[0053] Introduce a dynamic fuzzy neural network, adjust the membership function parameters of the input and output variables of the fuzzy logic controller according to the real-time spraying effect feedback, calculate the loss function, and the loss function includes the mean square error between the predicted quality and the actual quality, and the L2 regularization term of the difference before and after the adjustment of the membership function parameters; based on the loss function, calculate the gradient through the backpropagation algorithm, and update the membership function parameters of the input and output variables of the fuzzy logic controller in combination with the stochastic gradient descent method; repeat the steps of calculating the loss function and updating the parameters until the preset number of iterations is reached to obtain the sprayed part;

[0054] Use a high-precision three-dimensional laser scanner to perform an all-round scan on the sprayed part to generate a high-precision three-dimensional model;

[0055] Based on the high-precision three-dimensional model, adopt a multi-scale spraying quality assessment method to calculate the comprehensive quality score, integrate the comprehensive quality score, environmental parameters, part characteristics and spraying parameters into a full-cycle comprehensive data set, and store the full-cycle comprehensive data set in a time series database;

[0056] Based on the full-cycle comprehensive data set, extract the spraying quality influencing variables through the LightGBM algorithm, use the SHAP value to explain the LightGBM model and quantify the importance of each feature to obtain the ranking of the key factors affecting the spraying quality;

[0057] Update the network parameters of the multi-modal fusion neural network based on the spray quality impact variables. The multi-modal fusion neural network includes a one-dimensional convolutional neural network for processing time-series spray parameter data, a two-dimensional convolutional neural network for processing component geometric feature images, a long short-term memory network for processing environmental parameter time series, and a multi-layer perceptron subnet for processing static features. The outputs of each subnet are fused through an attention layer based on Transformer.

[0058] Update the algorithm parameters of the reinforcement learning algorithm based on the spray quality impact variables. Apply the updated multi-modal fusion neural network and the reinforcement learning algorithm to the next batch of automotive component spraying processes to continuously adjust the spray quality.

[0059] In the second aspect of the embodiments of the present invention,

[0060] Provide an automotive component spraying control system, including:

[0061] A first unit for real-time collecting spraying environment parameters through a temperature and humidity sensor, a pressure sensor, and an air quality sensor; obtaining the surface spectral information and three-dimensional structure data of automotive components by using a multi-spectral camera and a three-dimensional structured light scanner; inputting the spraying environment parameters, surface spectral information, and three-dimensional structure data into a pre-trained multi-modal fusion neural network to generate a comprehensive digital model containing environmental parameters and component features, and based on the comprehensive digital model, using a clustering partition algorithm to intelligently partition the surface of automotive components into multiple sub-regions;

[0062] A second unit for generating optimal spray parameters for each sub-region based on each sub-region and combining the current environmental parameters by using an intelligent agent trained by a reinforcement learning algorithm, optimizing the global spraying order by using an ant colony path planning algorithm based on the spatial relationship between the optimal spray parameters and the sub-regions to generate an optimal spray path; sending the optimal spray parameters and the optimal spray path to a spraying robot with an adaptive control function, and during the spraying process, real-time collecting spray state data through a high-speed camera and a coating thickness sensor, and inputting the spray state data into a pre-trained convolutional neural network to generate a real-time quality evaluation result of the spray;

[0063] The third unit is used to adjust the spraying parameters according to the deviation between the real-time quality assessment result and the preset quality target by using a pre-built fuzzy logic controller until the spraying of the current automotive part is completed, obtaining the sprayed part. Then, the sprayed part is scanned and detected to obtain the final spraying quality data. The final spraying quality data, environmental parameters, part characteristics, and spraying parameters are integrated into a full-cycle comprehensive data set. Based on the full-cycle comprehensive data set, the spraying quality influencing variables are extracted through the gradient boosting decision tree algorithm. Based on the spraying quality influencing variables, the network parameters of the multi-modal fusion neural network and the algorithm parameters of the reinforcement learning algorithm are updated.

[0064] In the third aspect of the embodiments of the present invention,

[0065] A kind of electronic device is provided, including:

[0066] A processor;

[0067] A memory for storing instructions executable by the processor;

[0068] Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0069] In the fourth aspect of the embodiments of the present invention,

[0070] A computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0071] The present invention realizes the comprehensive perception of the spraying environment and the object to be sprayed by collecting environmental parameters in real time through multiple sensors and combining a multi-spectral camera and a three-dimensional structured light scanner to obtain the surface information of the part. A multi-modal fusion neural network is used to generate a comprehensive digital model, and a clustering partition algorithm is used for intelligent partitioning to formulate personalized spraying strategies for different regions. This refined spraying method significantly improves the quality and accuracy of spraying, can adapt to automotive parts of different shapes and materials, and realizes a uniform and high-quality coating effect.

[0072] The present invention uses the reinforcement learning algorithm to generate the optimal spraying parameters for each sub-region and uses the ant colony path planning algorithm to optimize the global spraying sequence. This intelligent parameter setting and path planning method greatly improves the spraying efficiency, reduces unnecessary spraying time and paint waste. At the same time, through real-time quality assessment and dynamic adjustment of the fuzzy logic controller, the stability and consistency of the spraying process are ensured, the resource utilization is further optimized, and the production cost is reduced.

[0073] The present invention constructs a closed-loop adaptive optimization system. By collecting spraying state data in real time and combining it with the final spraying quality data, a comprehensive dataset for the entire cycle is generated. The gradient boosting decision tree algorithm is used to extract the variables affecting spraying quality, and based on this, the parameters of the multi-modal fusion neural network and the reinforcement learning algorithm are updated. This mechanism enables the system to continuously learn and optimize, adapt to different spraying environments and component characteristics, and achieve continuous improvement. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 is a schematic flowchart of the method for controlling the spraying of automotive components according to an embodiment of the present invention;

[0075] Figure 2 is a schematic structural diagram of the control system for spraying automotive components according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0076] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0077] The technical solutions of the present invention will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0078] Figure 1 is a schematic flowchart of the method for controlling the spraying of automotive components according to an embodiment of the present invention, as Figure 1 shown, the method includes:

[0079] S101. Real-time collect spraying environment parameters through a temperature and humidity sensor, a pressure sensor, and an air quality sensor; obtain the surface spectral information and three-dimensional structure data of the automotive component using a multi-spectral camera and a three-dimensional structured light scanner; input the spraying environment parameters, surface spectral information, and three-dimensional structure data into a pre-trained multi-modal fusion neural network to generate a comprehensive digital model containing environment parameters and component characteristics, and based on the comprehensive digital model, use a clustering partition algorithm to intelligently partition the surface of the automotive component into multiple sub-regions;

[0080] S102. Based on each sub-region, combined with the current environmental parameters, an intelligent agent trained by a reinforcement learning algorithm generates optimal spraying parameters for each sub-region. Based on the optimal spraying parameters and the spatial relationship between sub-regions, the ant colony path planning algorithm is used to optimize the global spraying order and generate an optimal spraying path; the optimal spraying parameters and the optimal spraying path are sent to a spraying robot with an adaptive control function. During the spraying process, the spraying state data is collected in real time through a high-speed camera and a coating thickness sensor, and the spraying state data is input into a pre-trained convolutional neural network to generate a real-time quality evaluation result of the spraying;

[0081] S103. An initially constructed fuzzy logic controller is adopted. According to the deviation between the real-time quality evaluation result and the preset quality target, the spraying parameters are adjusted until the current automotive part spraying is completed to obtain a sprayed part. The sprayed part is scanned and detected to obtain the final spraying quality data. The final spraying quality data, environmental parameters, part features, and spraying parameters are integrated into a full-cycle comprehensive data set. Based on the full-cycle comprehensive data set, the spraying quality influencing variables are extracted through a gradient boosting decision tree algorithm. Based on the spraying quality influencing variables, the network parameters of the multi-modal fusion neural network and the algorithm parameters of the reinforcement learning algorithm are updated.

[0082] The specific implementation method is as follows:

[0083] The automotive part spraying control method first collects the spraying environmental parameters in real time through multiple sensors. Specifically, a DHT22 temperature and humidity sensor is used to collect the environmental temperature and relative humidity, and the sampling frequency is 1Hz; a BMP280 barometric pressure sensor is used to collect the atmospheric pressure, and the sampling frequency is 0.1Hz; a PMS5003 air quality sensor is used to collect the PM2.5 and PM10 concentrations, and the sampling frequency is 0.2Hz. These environmental parameters have an important impact on the spraying quality and need to be monitored in real time.

[0084] At the same time, the surface feature information of the automotive part is obtained by using a multi-spectral camera and a three-dimensional structured light scanner. The multi-spectral camera adopts the model FLIR Blackfly S BFS-U3-51S5C-C, which can simultaneously collect images in the visible light, near-infrared, and short-wave infrared bands, with a resolution of 2448x2048 pixels and a frame rate of 30fps. The three-dimensional structured light scanner adopts the model Artec Leo, with a scanning accuracy of up to 0.1mm and a scanning speed of up to 80 frames per second. Through these two devices, the spectral information and three-dimensional structure data of the part surface can be comprehensively obtained.

[0085] Input the collected environmental parameters, surface spectral information, and three-dimensional structure data into a pre-trained multi-modal fusion neural network. This network uses ResNet50 as the backbone network and realizes the effective fusion of multi-modal features through an attention mechanism. The network outputs a 300-dimensional feature vector as the comprehensive digital model of the component. Based on this model, the DBSCAN clustering algorithm is used to intelligently partition the surface of the component. The eps parameter of the DBSCAN algorithm is set to 2mm, and the minPts parameter is set to 10. By adjusting these two parameters, different granularity partitioning results can be obtained. Generally, a medium-sized automotive component will be divided into 20 - 30 sub-regions.

[0086] For each sub-region, combined with the current environmental parameters, an intelligent agent trained by the Deep Deterministic Policy Gradient (DDPG) algorithm generates optimal spraying parameters. Both the Actor network and the Critic network of the DDPG algorithm adopt a three-layer fully connected network structure, and the number of neurons in the hidden layers is 256 and 128 respectively. The input is the sub-region features and environmental parameters, and the output is key parameters such as spraying distance, spraying speed, spraying angle, and atomization pressure. The reward function is designed as a weighted sum of the spraying quality index and energy consumption. Through repeated training, the intelligent agent can generate optimal spraying parameters for sub-regions with different features.

[0087] After obtaining the optimal spraying parameters for each sub-region, it is necessary to further optimize the global spraying sequence. Here, the ant colony algorithm is used for spraying path planning. The number of ant colonies is set to 50, the pheromone importance factor α = 1, the heuristic factor β = 2, and the pheromone evaporation coefficient ρ = 0.1. Through iterative optimization, an optimal spraying path with the shortest total path length and the highest spraying efficiency can be obtained.

[0088] Send the optimal spraying parameters and the optimal spraying path to an ABB IRB 5500 spraying robot with adaptive control function. This robot has 6 degrees of freedom, a repeat positioning accuracy of ±0.15mm, and a maximum load of 13kg, which is very suitable for spraying automotive components. During the spraying process, real-time spraying status data is collected through a Phantom v2640 high-speed camera (with a maximum frame rate of up to 25,030fps) and a Micro-Epsilon optoCONTROL 2520 coating thickness sensor (measurement range 46 - 650μm, resolution 1μm).

[0089] Input the collected spraying status data into a pre-trained ResNet18 convolutional neural network to generate a real-time quality assessment result of the spraying. This network is pre-trained on a large amount of historical spraying data and can quickly and accurately evaluate the current spraying quality. The evaluation results include multiple indicators such as coating uniformity, adhesion, and gloss, represented by scores from 0 to 100.

[0090] An pre-built fuzzy logic controller is adopted to adjust the spraying parameters according to the deviation between the real-time quality assessment result and the preset quality target. The input variables of the fuzzy controller are the quality deviation and the deviation change rate, and the output variables are the adjustment amount of the spraying distance, the adjustment amount of the spraying speed, etc. The fuzzy rule base contains 25 IF-THEN rules. Through fuzzy inference, an appropriate parameter adjustment amount can be obtained to achieve the closed-loop control of the spraying process.

[0091] After the spraying of the current automotive parts is completed, a Zeiss COMET L3D 2 3D scanner is used to scan and detect the sprayed parts to obtain the final spraying quality data. The measurement range of this scanner is 45 - 500 mm, and the accuracy can reach 2 μm. Detailed quality data such as the coating thickness distribution and surface roughness can be obtained through scanning.

[0092] The final spraying quality data, environmental parameters, part characteristics, and spraying parameters are integrated into a full-cycle comprehensive data set. Based on this data set, the variables affecting the spraying quality are extracted through the XGBoost gradient boosting decision tree algorithm. The main parameter settings of the XGBoost algorithm are as follows: the maximum tree depth is 6, the minimum number of samples in the leaf nodes is 5, the learning rate is 0.1, and the subsampling ratio is 0.8. Through feature importance analysis, the key variables affecting the spraying quality and their importance levels can be obtained.

[0093] Finally, based on the extracted variables affecting the spraying quality, the network parameters of the multi-modal fusion neural network and the algorithm parameters of the reinforcement learning algorithm are updated. For the neural network, the Adam optimizer is used for fine-tuning, and the learning rate is set to 0.0001; for the reinforcement learning algorithm, the exploration rate ε and the discount factor γ are adjusted. Through continuous model updates, the intelligent level and control accuracy of the system can be continuously improved.

[0094] A specific data case is as follows: For an automotive bumper, the environmental temperature is 25°C, the relative humidity is 60%, and the atmospheric pressure is 101.3 kPa. Through multi-modal feature extraction, a 300-dimensional digital model is obtained. After DBSCAN clustering, the bumper is divided into 25 sub-regions. For the central region, the optimal spraying parameters generated by the DDPG algorithm are: spraying distance 200 mm, spraying speed 300 mm / s, spraying angle 75°, and atomization pressure 0.3 MPa. The total length of the optimal path planned by the ant colony algorithm is 15.6 m. During the spraying process, the real-time quality assessment score is 92 points. The final scanning and detection results show that the average coating thickness is 52 μm, the thickness uniformity is ±3 μm, and the surface roughness Ra is 0.8 μm. The XGBoost algorithm analysis shows that the spraying distance and atomization pressure are the two most critical factors affecting the quality.

[0095] In an alternative embodiment, spray environment parameters, surface spectral information, and three-dimensional structure data are input into a pre-trained multi-modal fusion neural network to generate a comprehensive digital model that includes environmental parameters and component features. Based on the comprehensive digital model, an intelligent partitioning algorithm is used to intelligently partition the surface of automotive components into multiple sub-regions, including:

[0096] Collect spray environment parameters, where the spray environment parameters include temperature data, humidity data, air pressure data, and airborne particulate matter concentration data; collect hyperspectral information on the surface of automotive components, where the hyperspectral information includes spectral reflectance data for multiple bands; collect three-dimensional structure data of automotive components, where the three-dimensional structure data includes point cloud data.

[0097] Construct a multi-modal fusion neural network, where the multi-modal fusion neural network includes an environmental parameter branch, a spectral information branch, and a three-dimensional structure branch.

[0098] Input the spray environment parameters into the environmental parameter branch and use a one-dimensional convolutional neural network to process the time-series environmental data to obtain an environmental feature vector; input the hyperspectral information into the spectral information branch and use a three-dimensional convolutional neural network to extract spatial-spectral features to obtain a spectral feature vector; input the three-dimensional structure data into the three-dimensional structure branch and use a point cloud convolutional network to process the point cloud data to obtain a structural feature vector.

[0099] Use a multi-head self-attention module to fuse the environmental feature vector, spectral feature vector, and structural feature vector to obtain a fused feature; input the fused feature into a multi-layer perceptron and output a fixed-dimensional feature vector as the comprehensive digital model.

[0100] Based on the comprehensive digital model, use a clustering partitioning algorithm to intelligently partition the surface of automotive components and determine multiple sub-regions.

[0101] To implement an intelligent spraying method for automotive components based on a multi-modal fusion neural network and clustering partitioning, the following steps can be specifically carried out:

[0102] First, collect spray environment parameters. Use a temperature and humidity sensor to collect temperature and humidity data with a sampling frequency of 1 Hz and continuously collect for 24 hours to obtain 86,400 temperature data points and 86,400 humidity data points. Use a barometric pressure sensor to collect barometric pressure data with a sampling frequency of 0.1 Hz and continuously collect for 24 hours to obtain 8,640 barometric pressure data points. Use a particulate matter concentration detector to collect PM2.5 and PM10 concentration data with a sampling frequency of 0.01 Hz and continuously collect for 24 hours to obtain 864 PM2.5 concentration data points and 864 PM10 concentration data points.

[0103] Secondly, collect the hyperspectral information of the surface of automotive parts. Use a hyperspectral imager to scan the surface of the parts to obtain the reflectance data of 210 bands in the wavelength range of 400 - 2500 nm. The scanning resolution is 1 mm × 1 mm. For a 1 m × 1 m part surface, a hyperspectral data cube of 1000 × 1000 × 210 can be obtained.

[0104] Then, collect the three-dimensional structure data of automotive parts. Use a three-dimensional laser scanner to scan the parts omni-directionally to obtain point cloud data. The scanning accuracy is 0.1 mm. For a part of 1 m × 1 m × 0.5 m, point cloud data of about 5 million points can be obtained, and each point contains spatial coordinate (x, y, z) information.

[0105] Next, construct a multi-modal fusion neural network. The environmental parameter branch uses 5 one-dimensional convolutional layers with a convolutional kernel size of 3, a stride of 1, and the number of channels being 32, 64, 128, 256, and 512 in sequence. The spectral information branch uses 4 three-dimensional convolutional layers with a convolutional kernel size of 3 × 3 × 3, a stride of 1, and the number of channels being 32, 64, 128, and 256 in sequence. The three-dimensional structure branch uses the PointNet++ network, which includes 4 set abstraction layers and 2 feature propagation layers. The multi-head self-attention module uses 8 attention heads with a hidden layer dimension of 512. The multi-layer perceptron contains 3 fully connected layers with hidden layer dimensions of 1024, 512, and 256 in sequence.

[0106] Input the collected data into the constructed neural network for processing. The environmental parameter branch inputs an 86400 × 5 environmental parameter matrix, and after one-dimensional convolutional processing, a 512-dimensional environmental feature vector is obtained. The spectral information branch inputs a 1000 × 1000 × 210 hyperspectral data cube, and after three-dimensional convolutional processing, a 256-dimensional spectral feature vector is obtained. The three-dimensional structure branch inputs the point cloud data of 5 million points, and after processing by the PointNet++ network, a 128-dimensional structure feature vector is obtained.

[0107] Then, use the multi-head self-attention module to fuse the three feature vectors. First, concatenate the three feature vectors into an 896-dimensional vector, and then input it into 8 parallel self-attention sub-modules. Each sub-module generates query vectors, key vectors, and value vectors through linear transformation, calculates the attention weights, and sums them up with weights to obtain a 64-dimensional output. The outputs of the 8 sub-modules are concatenated to obtain a 512-dimensional fusion feature.

[0108] Finally, input the fusion feature into a 3-layer fully connected network to obtain a 256-dimensional comprehensive digital model feature vector. This feature vector contains the comprehensive information of environmental parameters, surface spectra, and three-dimensional structures.

[0109] Based on the generated comprehensive digital model, the K-means clustering algorithm is used to intelligently partition the surface of parts. First, the 256-dimensional feature vector is reduced to 3 dimensions, and then the number of clustering centers K is set to 5, and the reduced feature points are clustered. The clustering result is the 5 sub-regions on the surface of the parts, and different regions have different environmental adaptabilities and surface characteristics.

[0110] Through the above steps, an intelligent spraying partition method for automotive parts based on multi-modal fusion and clustering partition is realized. This method comprehensively considers environmental factors, surface characteristics and structural features, and can more accurately partition the surface of parts, providing a basis for subsequent intelligent spraying.

[0111] In an alternative embodiment, the clustering partition algorithm includes:

[0112] Construct a clustering partition algorithm based on K-means clustering; initialize the clustering center, and use the feature vectors in the comprehensive digital model as the initial data points;

[0113] Calculate the distance from each data point to each clustering center. Introduce spatial constraints during distance calculation, add spatial coordinate terms to the distance calculation formula, and assign adjustable weight coefficients; assign each data point to the nearest clustering center; calculate the average value of all data points in each cluster, determine the new clustering center, and update; calculate the local curvature of each data point, and use the local curvature as the clustering feature dimension;

[0114] Repeat the iteration until the change in the position of the clustering center is less than the preset clustering center change threshold to obtain the optimal partition scheme;

[0115] Calculate the silhouette coefficient of the optimal partition scheme through a similarity measurement method;

[0116] Within the preset range of the number of partitions, repeat the initialization and iterative calculation processes to obtain the silhouette coefficients under different numbers of partitions, and select the partition scheme with the largest silhouette coefficient as the final intelligent partition result.

[0117] In the specific implementation, a clustering partition algorithm is constructed based on K-means clustering to intelligently partition the comprehensive digital model. First, initialize the clustering center, and use the feature vectors in the comprehensive digital model as the initial data points. The feature vectors contain information such as geometric features and material properties.

[0118] Next, calculate the distance from each data point to each clustering center. Introduce spatial constraints during distance calculation, add spatial coordinate terms to the distance calculation formula, and assign adjustable weight coefficients. For example, the Euclidean distance weighted formula can be used, and the weight of the spatial coordinate term can be set to 0.3 - 0.5. This can ensure that points with similar spatial positions are more likely to be divided into the same cluster.

[0119] Then each data point is assigned to the nearest cluster center. Calculate the average value of all data points in each cluster, determine the new cluster center, and update the position of the cluster center.

[0120] During the iteration process, the local curvature of each data point is also calculated and used as a feature dimension for clustering. The local curvature can be calculated by fitting a quadratic surface around the data points. This helps to identify feature regions in the model, such as edges, corners, etc.

[0121] Repeat the process of iterative calculation and updating of the cluster center until the change in the position of the cluster center is less than a preset threshold, such as 0.001. In this way, the optimal partitioning scheme for the current number of partitions is obtained.

[0122] To evaluate the partitioning effect, the silhouette coefficient of the optimal partitioning scheme is calculated. The silhouette coefficient measures the compactness and separation of the clusters, and its value ranges from -1 to 1. The closer it is to 1, the better the clustering effect.

[0123] Within a preset range of the number of partitions, such as 2 to 10, repeat the initialization and iterative calculation processes to obtain the silhouette coefficients for different numbers of partitions. Select the partitioning scheme with the largest silhouette coefficient as the final intelligent partitioning result.

[0124] For example, for a car front bumper model, the maximum silhouette coefficient of 0.72 may be obtained when the number of partitions is 5. This indicates that dividing the model into 5 regions is the most appropriate, ensuring both the similarity within each region and the difference between regions.

[0125] The final intelligent partitioning result is visually presented, with different regions marked in different colors. This provides a basis for subsequent finite element analysis, 3D printing, etc. By adjusting the weight coefficients in distance calculation, changing clustering features, etc., the partitioning effect can be further optimized to better meet the requirements of engineering applications.

[0126] In an alternative embodiment, based on each sub-region, combined with the current environmental parameters, an intelligent agent trained using a reinforcement learning algorithm generates optimal spraying parameters for each sub-region, including:

[0127] Construct a main network based on a deep Q-network. The main network includes an input layer, two hidden layers, and an output layer. The number of nodes in the input layer is the same as the dimension of the state vector, the number of nodes in the output layer is equal to the size of the action space, the hidden layers use the ReLU activation function, and the output layer uses a linear activation function;

[0128] Determine the state space based on the geometric features, material properties, and environmental parameters of the sub-region; determine the action space based on the spraying parameters; calculate the weighted sum of the coating quality indicators in the sub-region based on the preset coating quality indicators, and construct a reward function;

[0129] Construct a simulation environment that simulates the spraying process of the sub-region on the surface of automotive parts based on the state space, action space, and reward function;

[0130] Construct an experience pool for storing the experience data obtained from the interaction between the intelligent agent and the simulation environment, initialize the experience pool, create a target network based on the structure of the main network, and calculate the target Q value;

[0131] Adopt the ε-greedy method, set the exploration strategy, initially set the ε value to 0.9, and linearly decay it to 0.1 as the number of training rounds increases;

[0132] The intelligent agent selects spraying parameters according to the current sub-region state and the exploration strategy, interacts with the simulation environment, obtains the reward and the next sub-region state, and calculates the temporal difference error based on the current sub-region state, the selected spraying parameters, the obtained reward, the next sub-region state, and the target Q value calculated by the target network;

[0133] Take the current sub-region state, the selected spraying parameters, the obtained reward, the next sub-region state, and the temporal difference error as an experience data, store and update the experience pool; assign a sampling probability to each experience data based on the temporal difference error in the experience pool, select batch data based on the sampling probability, and calculate the loss function based on the batch data, the main network, and the target network;

[0134] Minimize the loss function through the gradient descent method, update the parameters of the main network, assign the weighted average value of the main network parameters to the corresponding parameters of the target network, update the parameters of the target network, add Gaussian noise to the parameters of the main network and the target network, and continuously iterate until the preset number of training rounds is reached, determine the trained main network, input the current state of each sub-region into the trained main network, and input the optimal spraying parameter combination of the sub-region.

[0135] To optimize the surface spraying parameters of automotive parts based on deep reinforcement learning, it is first necessary to construct a deep Q-network as the main network. This main network consists of an input layer, two hidden layers, and an output layer. The number of nodes in the input layer is the same as the dimension of the state vector. For example, it can be set to 10 nodes, corresponding to the geometric features (such as area, curvature, etc.), material properties (such as surface roughness, hardness, etc.), and environmental parameters (such as temperature, humidity, etc.) of the sub-region respectively. The two hidden layers are set with 64 and 32 nodes respectively, and the ReLU activation function is used. The number of nodes in the output layer is equal to the size of the action space. For example, it can be set to 5 nodes, corresponding to spraying parameters such as spray gun distance, spraying angle, spraying speed, paint flow rate, and atomization pressure. The output layer uses a linear activation function.

[0136] Next, determine the state space based on the features of the sub-region. For example, the area, curvature, surface roughness, hardness, environmental temperature, etc. of the sub-region can be used as state variables to form a 10-dimensional state vector. At the same time, determine a 5-dimensional action space based on the spraying parameters, including the aforementioned 5 spraying parameters. Then, based on the preset coating quality indicators, such as coating thickness uniformity, adhesion, etc., calculate the weighted sum of these indicators in the sub-region to construct a reward function. For example, the weight of coating thickness uniformity can be set to 0.6, and the weight of adhesion can be set to 0.4, and the weighted sum is calculated as the reward value.

[0137] Based on the state space, action space, and reward function defined above, construct a simulation environment that simulates the spraying process of the sub-region on the surface of automotive parts. This simulation environment can simulate the spraying process and calculate the corresponding reward value according to the input state and action.

[0138] Subsequently, construct an experience pool to store the experience data obtained from the interaction between the intelligent agent and the simulation environment. The capacity of the experience pool can be set to 10000 and randomly initialized. Based on the structure of the main network, create a target network with the same structure to calculate the target Q value.

[0139] During the training process, use the ε-greedy method to set the exploration strategy. Initially, set the ε value to 0.9, and as the number of training rounds increases, the ε value linearly decays to 0.1. For example, this decay process can be achieved in 10000 rounds of training.

[0140] The intelligent agent selects spraying parameters according to the current sub-region state and exploration strategy. Specifically, generate a random number between 0 and 1. If it is less than the current ε value, randomly select an action; otherwise, select the action with the largest Q value. The intelligent agent interacts with the simulation environment to obtain the reward and the next sub-region state. Based on the current sub-region state, the selected spraying parameters, the obtained reward, the next sub-region state, and the target Q value calculated by the target network, calculate the temporal difference error.

[0141] Take the current sub-region state, selected spraying parameters, obtained rewards, next sub-region state, and temporal difference error in the above interaction process as an experience data, store it in and update the experience pool. Allocate sampling probabilities for each experience data based on the temporal difference error in the experience pool. The larger the temporal difference error, the higher the sampling probability. For example, the temporal difference error can be normalized as the sampling probability. Based on these sampling probabilities, select a batch of data from the experience pool, and the batch size can be set to 64. Use this batch of data, the main network, and the target network to calculate the loss function.

[0142] Minimize the loss function through gradient descent to update the main network parameters. The Adam optimizer can be used, and the learning rate is set to 0.001. Assign the weighted average of the updated main network parameters to the corresponding parameters of the target network to achieve soft update of the target network. The update frequency can be set to update once every 10 training rounds, and the update coefficient is set to 0.01. To enhance the robustness of the model, Gaussian noise with a mean of 0 and a standard deviation of 0.01 can be added to the main network parameters and the target network parameters.

[0143] Continuously iterate the above process until the preset number of training rounds is reached, such as 10,000 rounds. After determining the main network with training completed, input the current state of each sub-region into this network, and the optimal spraying parameter combination for this sub-region can be obtained. For example, for a sub-region with an area of 100 square centimeters, a curvature of 0.05, a surface roughness of 3μm, a hardness of 60HRC, and an environmental temperature of 25°C, the optimal spraying parameters output by the trained network may be: spray gun distance 20cm, spraying angle 75°, spraying speed 30cm / s, paint flow rate 100ml / min, atomization pressure 0.4MPa.

[0144] In this way, customized optimal spraying parameters can be generated for each sub-region on the surface of automotive parts, thereby optimizing the overall spraying effect.

[0145] In an alternative embodiment, based on the spatial relationship between the optimal spraying parameters and the sub-regions, the ant colony path planning algorithm is used to optimize the global spraying order, and the steps to generate the optimal spraying path include:

[0146] Based on each sub-region, generate a corresponding node. Based on the spatial relationship between the sub-regions, determine the edges between the nodes. Based on the nodes and edges, determine the graph structure;

[0147] Calculate the edge weights based on the spatial distance between sub-regions, the optimal spraying parameters corresponding to each sub-region, and the surface curvature change of the sub-regions. Among them, determine the spatial distance based on the Euclidean distance between adjacent sub-regions, calculate the similarity degree between the corresponding vectors of the optimal spraying parameters of two sub-regions through cosine similarity to determine the spraying parameter similarity, calculate the curvature between adjacent sub-regions to determine the curvature difference, and obtain the surface curvature change.

[0148] Initialize the ant population and the corresponding pheromone matrix according to the preset number of ant populations. Each ant randomly selects a sub-region as the starting sub-region.

[0149] Determine the heuristic information based on the edge weights, determine the pheromone concentration based on the pheromone matrix, calculate the state transition probability according to the heuristic information and the pheromone concentration, determine the probability of each ant selecting the next sub-region, select the sub-region corresponding to the maximum selection probability, and move the ant to the corresponding sub-region.

[0150] After the ant moves, update the local pheromone of the edge passed during the movement. The local pheromone is updated based on the corresponding pheromone concentration in the pheromone matrix, the preset pheromone evaporation coefficient, and the preset initial pheromone concentration.

[0151] Repeat the steps of sub-region selection and local pheromone update until all ants complete a full path. Based on the spatial distance between sub-regions and the time distance of spray gun adjustment, determine the path length of the ants. Based on the spraying parameter similarity and curvature change between adjacent sub-regions, determine the spraying quality score corresponding to the ants. Update the global pheromone based on the path length and the spraying quality score. Repeat until the preset maximum number of iterations is reached, select the spraying path corresponding to the maximum spraying quality score, and determine the optimal spraying path.

[0152] To achieve the optimal spraying path planning based on the ant colony algorithm, first, it is necessary to divide the spraying area to obtain multiple sub-regions. For each sub-region, generate a corresponding node, and determine the connecting edges between the nodes based on the spatial relationship between the sub-regions, thereby constructing a graph structure. This graph structure will serve as the basis for the movement of ants.

[0153] Next, it is necessary to calculate the weights of each edge in the graph. The edge weights are determined by three factors: spatial distance, spraying parameter similarity, and surface curvature change. The spatial distance is obtained by calculating the Euclidean distance between the center points of adjacent sub-regions. The spraying parameter similarity is obtained by calculating the cosine similarity between the optimal spraying parameter vectors of two sub-regions. The higher the similarity, the smaller the weight. The surface curvature change is represented by calculating the difference in the average curvature of adjacent sub-regions. The greater the curvature difference, the greater the weight. After these three factors are normalized, they are weighted and summed in a certain proportion to obtain the final edge weight.

[0154] When initializing the ant colony, the number of ants needs to be set, such as 50 ants. Each ant randomly selects a starting sub-region. At the same time, initialize the pheromone matrix. The size of the matrix is the square of the number of nodes, and the initial value can be set to a small constant, such as 0.1.

[0155] During the movement of the ants, each ant calculates the selection probability of the adjacent nodes that it may move to next based on the current node it is at. The selection probability consists of two parts: heuristic information and pheromone concentration. The heuristic information reflects the attractiveness of the edge and can be taken as the reciprocal of the edge weight. The pheromone concentration is obtained from the pheromone matrix. These two parts of information are weighted according to a certain ratio (such as 50% each) to obtain the normalized selection probability. The ant randomly selects the next node to move according to this probability distribution.

[0156] Whenever an ant completes a move, it is necessary to update the local pheromone of the passed edge. The update formula is: new pheromone = (1 - ρ) * old pheromone + ρ * initial pheromone. Where ρ is the pheromone evaporation coefficient, which can be taken as 0.1. This local update is beneficial to increasing the diversity of the paths.

[0157] When all ants have completed a complete path, it is necessary to evaluate the path quality of each ant and update the global pheromone. The path quality consists of two parts: path length and spraying quality score. The path length takes into account the spatial distance between sub-regions and the spray gun adjustment time. The spraying quality score is based on the similarity of spraying parameters and the curvature change between adjacent sub-regions on the path. The shorter the path and the higher the spraying quality score, the better the path quality. Update the global pheromone according to the path quality. The better the path quality, the greater the increase in pheromone.

[0158] Repeat the above process until the preset maximum number of iterations is reached, such as 1000 times. Finally, select the path with the highest spraying quality score among all iterations as the optimal spraying path.

[0159] The following uses a specific case to illustrate the implementation process of this method:

[0160] Suppose the outer surface of a car body is divided into 20 sub-regions. First, construct a graph structure with 20 nodes, and there are edge connections between adjacent sub-regions. In the calculated edge weight matrix, for example, the edge weight from node 1 to node 2 is 0.8, and the edge weight from node 1 to node 3 is 1.2, etc.

[0161] Initialize 50 ants, randomly distributed on 20 nodes. The pheromone matrix is initialized to 0.1.

[0162] In the first iteration, assume the movement path of an ant is: 1->2->5->8->12->15->18->20->17->14->11->7->4->3->6->9->10->13->16->19.

[0163] During the movement, the local pheromone is continuously updated. For example, after the ant moves from node 1 to node 2, the pheromone of edge (1, 2) is updated to: 0.9*0.1 + 0.1*0.1 = 0.09.

[0164] After a complete iteration, calculate that the total length of the ant's path is 100 (unit: cm), and the spraying quality score is 85 (full score 100). Accordingly, the global pheromone is updated, such as adding 0.05 pheromone to all edges on this path.

[0165] After 1000 iterations, the optimal path obtained is: 1->4->7->10->13->16->19->20->17->14->11->8->5->2->3->6->9->12->15->18. The total length of this path is 90 cm, and the spraying quality score is 92, which is the best result among all iterations.

[0166] This method based on the ant colony algorithm can effectively balance the spraying path length and the spraying quality, obtain the optimal spraying path that takes into account both efficiency and effect, thereby improving the overall performance of the automatic spraying system.

[0167] In an alternative embodiment, the path length has the following formula:

[0168] ;

[0169] where L represents the total path length, m represents the number of sub-regions included in the path, i represents the index of the sub-region in the path, r(·) represents the sub-region corresponding to the index in the path, d(·, ·) represents the spatial distance between sub-regions, T represents the time coefficient required to adjust the spraying parameters, P(·) represents the optimal spraying parameter vector, and ||·|| represents calculating the Euclidean distance;

[0170] The spraying quality score has the following formula:

[0171] ;

[0172] where Q represents, λ represents the weight coefficient, sim(·) represents calculating the spraying parameter similarity, c(·) represents calculating the average curvature, and |·| represents the absolute value of the curvature difference.

[0173] In practical applications, this technical solution can be implemented in the following ways:

[0174] First, it is necessary to obtain the three-dimensional model data of the workpiece to be sprayed and divide it into multiple sub-regions. For each sub-region, calculate its geometric features such as surface curvature, etc. Then, according to the process requirements, set the optimal spraying parameters for each sub-region, including the spray gun distance, spraying angle, spraying speed, etc.

[0175] Next, determine the spraying path. Starting from the initial sub-region, traverse all sub-regions in sequence and calculate the spatial distance between adjacent sub-regions. At the same time, consider the time required to adjust the spraying parameters to obtain the total path length. Specifically, for the i-th and (i + 1)-th sub-regions in the path, calculate the distance between their center points, and then add the time to adjust the spraying parameters multiplied by the degree of parameter difference. Accumulate this "distance" between all adjacent sub-regions to obtain the total path length.

[0176] While calculating the path length, it is also necessary to evaluate the spraying quality. Two main factors are considered: one is the similarity of spraying parameters between adjacent sub-regions, and the other is the difference in average curvature between sub-regions. The higher the similarity of spraying parameters and the smaller the curvature difference, the better the spraying quality. These two indicators can be calculated separately and weighted and summed according to a certain weight to obtain the final spraying quality score.

[0177] For example, assume that the workpiece to be sprayed is divided into 5 sub-regions. The optimal spraying parameters for the first sub-region are (10, 30, 50), and for the second sub-region are (12, 35, 55). The distance between their center points is 100 mm, and it takes 2 seconds to adjust the parameters. Then the path length between these two sub-regions can be calculated as: 100 + 2 * ((12 - 10)^2 + (35 - 30)^2 + (55 - 50)^2)^0.5 ≈ 111.4 mm. Calculate other adjacent sub-regions in the same way and accumulate to obtain the total path length.

[0178] For the spraying quality score, assume that the average curvatures of the first and second sub-regions are 0.01 and 0.015 respectively, and the parameter similarity is 0.9. Take the weight coefficient λ = 0.6. Then the spraying quality score between these two sub-regions is: 0.6 * 0.9 + 0.4 * (1 - |0.01 - 0.015|) = 0.938. Calculate other adjacent sub-regions in the same way and take the average as the overall score.

[0179] Through the above method, the performance of a given spraying path can be comprehensively evaluated. In practical applications, multiple candidate paths can be generated by adjusting the sub-region division, changing the traversal order, etc., and the optimal path can be selected using this evaluation system, so as to achieve intelligent spraying path planning.

[0180] In an alternative embodiment, a pre - constructed fuzzy logic controller is adopted to adjust the spraying parameters according to the deviation between the real - time quality assessment result and the preset quality target until the spraying of the current automotive part is completed, obtaining the sprayed - completed part. The sprayed - completed part is scanned and detected to obtain the final spraying quality data. The final spraying quality data, environmental parameters, part characteristics, and spraying parameters are integrated into a full - cycle comprehensive data set. Based on the full - cycle comprehensive data set, through the gradient - boosting decision tree algorithm, the spraying quality influencing variables are extracted. Based on the spraying quality influencing variables, the network parameters of the multi - modal fusion neural network and the algorithm parameters of the reinforcement learning algorithm are updated, including:

[0181] According to the deviation between the real - time quality assessment result and the preset quality target, the spraying parameters are adjusted by a pre - constructed fuzzy logic controller until the spraying of the current automotive part is completed, including:

[0182] A dynamic fuzzy neural network is introduced. According to the real - time spraying effect feedback, the membership function parameters of the input and output variables of the fuzzy logic controller are adjusted, and the loss function is calculated. The loss function includes the mean square error between the predicted quality and the actual quality, and the L2 regularization term of the difference before and after the adjustment of the membership function parameters. Based on the loss function, the gradient is calculated through the back - propagation algorithm, and the membership function parameters of the input and output variables of the fuzzy logic controller are updated in combination with the stochastic gradient descent method. The steps of calculating the loss function and updating the parameters are repeated until the preset number of iterations is reached, obtaining the sprayed - completed part;

[0183] The sprayed - completed part is scanned in all directions using a high - precision 3D laser scanner to generate a high - precision 3D model;

[0184] Based on the high - precision 3D model, a multi - scale spraying quality assessment method is adopted to calculate the comprehensive quality score. The comprehensive quality score, environmental parameters, part characteristics, and spraying parameters are integrated into a full - cycle comprehensive data set, and the full - cycle comprehensive data set is stored in a time - series database;

[0185] Based on the full - cycle comprehensive data set, the spraying quality influencing variables are extracted through the LightGBM algorithm, and the LightGBM model is explained using SHAP values and the importance of each feature is quantified to obtain the ranking of the key factors affecting the spraying quality;

[0186] Update the network parameters of the multi-modal fusion neural network based on the spray quality influence variables. The multi-modal fusion neural network includes a one-dimensional convolutional neural network for processing time-series spray parameter data, a two-dimensional convolutional neural network for processing component geometric feature images, a long short-term memory network for processing environmental parameter time series, and a multi-layer perceptron subnet for processing static features. The outputs of each subnet are fused through an attention layer based on Transformer.

[0187] Update the algorithm parameters of the reinforcement learning algorithm based on the spray quality influence variables, and apply the updated multi-modal fusion neural network and the reinforcement learning algorithm to the spray process of the next batch of automotive components to continuously adjust the spray quality.

[0188] In practical applications, the specific implementation of the method of the present invention is as follows:

[0189] First, based on a pre-constructed fuzzy logic controller, adjust the spray parameters according to the deviation between the real-time quality evaluation result and the preset quality target. Specifically, introduce a dynamic fuzzy neural network to dynamically adjust the membership function parameters of the input and output variables of the fuzzy logic controller according to the real-time spray effect feedback. Calculate the loss function, including the mean square error between the predicted quality and the actual quality, and the L2 regularization term of the difference before and after the adjustment of the membership function parameters. For example, the weight of the mean square error between the predicted quality and the actual quality can be set to 0.7, and the weight of the L2 regularization term can be set to 0.3. Based on the loss function, calculate the gradient through the backpropagation algorithm, and update the membership function parameters of the input and output variables of the fuzzy logic controller in combination with the stochastic gradient descent method. Repeat the steps of calculating the loss function and updating the parameters until the preset number of iterations (such as 100 times) is reached to obtain the components after spraying.

[0190] Next, use a high-precision three-dimensional laser scanner to perform a full-range scan of the components after spraying to generate a high-precision three-dimensional model. The scanning accuracy can reach 0.1 mm, and the scanning frequency is 200 Hz. Based on the generated high-precision three-dimensional model, adopt a multi-scale spray quality evaluation method to calculate the comprehensive quality score. The multi-scale evaluation method includes three levels: macroscopic appearance evaluation, mesoscopic surface topography evaluation, and microscopic coating structure evaluation. The macroscopic appearance evaluation mainly considers indicators such as color difference and gloss; the mesoscopic surface topography evaluation mainly considers indicators such as surface roughness and waviness; the microscopic coating structure evaluation mainly considers indicators such as coating thickness uniformity and porosity. The evaluation results of each level are weighted according to a certain weight (such as 0.4:0.3:0.3) to obtain the final comprehensive quality score.

[0191] Integrate the comprehensive quality score, environmental parameters (such as temperature, humidity, dust concentration, etc.), component characteristics (such as material, shape complexity, etc.), and spraying parameters (such as spray gun movement speed, spraying distance, paint flow rate, etc.) into a full-cycle comprehensive dataset and store it in a time series database. The time series database uses InfluxDB, which can efficiently store and query time series data.

[0192] Based on the stored full-cycle comprehensive dataset, extract the variables affecting spraying quality through the LightGBM algorithm. LightGBM is an ensemble learning algorithm based on decision trees, with advantages such as fast training speed and low memory occupancy. In this embodiment, the main parameters of LightGBM are set as follows: the maximum depth of the tree is 6, the minimum data volume of the leaf nodes is 20, and the minimum gain of feature splitting is 0.1. Use SHAP (SHapley Additive exPlanations) values to explain the LightGBM model and quantify the importance of each feature, and obtain the ranking of key factors affecting spraying quality. For example, the possible ranking result is: spraying distance > paint flow rate > spray gun movement speed > environmental temperature > surface roughness of parts, etc.

[0193] According to the extracted variables affecting spraying quality, update the network parameters of the multi-modal fusion neural network. The multi-modal fusion neural network includes multiple sub-networks: a one-dimensional convolutional neural network (1D-CNN) for processing time series spraying parameter data, a two-dimensional convolutional neural network (2D-CNN) for processing component geometric feature images, a long short-term memory network (LSTM) for processing environmental parameter time series, and a multi-layer perceptron (MLP) sub-network for processing static features. The 1D-CNN adopts a 3-layer convolutional structure, and the convolutional kernel sizes are 3, 5, and 7 respectively, followed by a max-pooling layer after each layer; the 2D-CNN adopts a ResNet-18 architecture; the LSTM adopts a 2-layer structure, and each layer contains 128 hidden units; the MLP adopts a 3-layer structure, and the number of neurons in each layer is 64, 32, and 16 respectively. The outputs of each sub-network are fused through an attention layer based on Transformer, and the Transformer adopts a 4-head attention mechanism with a hidden layer dimension of 256.

[0194] At the same time, based on the extracted variables affecting spraying quality, update the algorithm parameters of the reinforcement learning algorithm. The reinforcement learning algorithm uses a deep Q-network (DQN), which includes 3 fully connected layers, and the number of neurons in each layer is 128, 64, and 32 respectively. Update the parameters of the epsilon-greedy strategy, set the initial epsilon value to 0.9, decay by 0.995 per round, and the minimum value is 0.01. Update the experience replay buffer size to 10000, and the size of each sampling batch is 64.

[0195] Finally, apply the updated multi-modal fusion neural network and reinforcement learning algorithm to the spraying process of the next batch of automotive parts, and continuously adjust the spraying quality. Specifically, the multi-modal fusion neural network is used to predict the spraying quality in the current state, and the reinforcement learning algorithm selects the optimal spraying parameter adjustment strategy according to the prediction result. Through continuous iterative optimization, continuous improvement of the spraying quality is achieved.

[0196] Through the above technical solutions, the method of the present invention can effectively improve the spraying quality of automotive parts, reduce the defective rate, and improve production efficiency. This method has strong adaptability and robustness, and can cope with different types of parts and changing production environments.

[0197] Figure 2 It is a schematic structural diagram of the spraying control system for automotive parts in an embodiment of the present invention, as Figure 2 shown, the system includes:

[0198] The first unit is used to collect spraying environment parameters in real time through temperature and humidity sensors, air pressure sensors, and air quality sensors; obtain the surface spectral information and three-dimensional structure data of automotive parts using a multi-spectral camera and a three-dimensional structured light scanner; input the spraying environment parameters, surface spectral information, and three-dimensional structure data into a pre-trained multi-modal fusion neural network to generate a comprehensive digital model containing environmental parameters and part characteristics, and based on the comprehensive digital model, use the clustering partition algorithm to intelligently partition the surface of automotive parts into multiple sub-regions;

[0199] The second unit is used to, based on each sub-region, combine the current environmental parameters, and use the intelligent agent trained by the reinforcement learning algorithm to generate the optimal spraying parameters for each sub-region. Based on the optimal spraying parameters and the spatial relationship between sub-regions, use the ant colony path planning algorithm to optimize the global spraying sequence and generate the optimal spraying path; send the optimal spraying parameters and the optimal spraying path to a spraying robot with an adaptive control function. During the spraying process, collect spraying state data in real time through a high-speed camera and a coating thickness sensor, and input the spraying state data into a pre-trained convolutional neural network to generate a real-time quality evaluation result of the spraying;

[0200] The third unit is used to adopt a pre-constructed fuzzy logic controller to adjust the spraying parameters according to the deviation between the real-time quality evaluation result and the preset quality target until the spraying of the current automotive part is completed to obtain the sprayed part. Scan and detect the sprayed part to obtain the final spraying quality data, integrate the final spraying quality data, environmental parameters, part characteristics, and spraying parameters into a full-cycle comprehensive data set, and based on the full-cycle comprehensive data set, extract the spraying quality influencing variables through the gradient boosting decision tree algorithm. Based on the spraying quality influencing variables, update the network parameters of the multi-modal fusion neural network and the algorithm parameters of the reinforcement learning algorithm.

[0201] In a third aspect of the embodiments of the present invention,

[0202] a kind of electronic device is provided, including:

[0203] a processor;

[0204] a memory for storing instructions executable by the processor;

[0205] wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0206] In a fourth aspect of the embodiments of the present invention,

[0207] a computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0208] The present invention may be a method, a device, a system, and / or a computer program product. The computer program product may include a computer-readable storage medium, on which computer-readable program instructions for executing various aspects of the present invention are loaded.

[0209] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for controlling the spraying of automobile parts, characterized in that: include: The spraying environment parameters are collected in real time through temperature and humidity sensors, air pressure sensors and air quality sensors; the surface spectrum information and three-dimensional structure data of automobile parts are obtained by using multi-spectral cameras and three-dimensional structured light scanners; the spraying environment parameters, surface spectrum information and three-dimensional structure data are input into the pre-trained multimodal fusion neural network to generate a comprehensive digital model containing environmental parameters and parts characteristics. Based on the comprehensive digital model, the surface of automobile parts is intelligently partitioned into multiple sub-areas using a clustering partitioning algorithm; Based on each sub-area and combined with the current environmental parameters, an intelligent agent trained by a reinforcement learning algorithm generates optimal spraying parameters for each sub-area. Based on the optimal spraying parameters and the spatial relationship between the sub-areas, an ant colony path planning algorithm is used to optimize the global spraying sequence and generate the optimal spraying path. The optimal spraying parameters and the optimal spraying path are sent to a spraying robot with an adaptive control function. During the spraying process, the spraying status data is collected in real time through a high-speed camera and a coating thickness sensor. The spraying status data is input into a pre-trained convolutional neural network to generate real-time quality evaluation results for the spraying. A pre-built fuzzy logic controller is used to adjust the spraying parameters according to the deviation between the real-time quality assessment result and the preset quality target until the spraying of the current automobile parts is completed, and the sprayed parts are obtained. The sprayed parts are scanned and inspected to obtain the final spraying quality data. The final spraying quality data, environmental parameters, component characteristics and spraying parameters are integrated into a full-cycle comprehensive data set. Based on the full-cycle comprehensive data set, the gradient boosting decision tree algorithm is used to extract the spraying quality influencing variables. Based on the spraying quality influencing variables, the network parameters of the multimodal fusion neural network and the algorithm parameters of the reinforcement learning algorithm are updated.

2. The method according to claim 1, characterized in that: The spraying environment parameters, surface spectrum information and three-dimensional structure data are input into the pre-trained multimodal fusion neural network to generate a comprehensive digital model containing environmental parameters and component characteristics. Based on the comprehensive digital model, the surface of the automotive parts is intelligently partitioned using a clustering partitioning algorithm and divided into multiple sub-areas, including: Collect spraying environment parameters, including temperature data, humidity data, air pressure data, and air particle concentration data; collect hyperspectral information on the surface of automobile parts, including spectral reflectance data of multiple bands; collect three-dimensional structure data of automobile parts, including point cloud data; Constructing a multimodal fusion neural network, wherein the multimodal fusion neural network includes an environmental parameter branch, a spectral information branch, and a three-dimensional structure branch; The spraying environment parameters are input into the environment parameter branch, and the time series environment data is processed using a one-dimensional convolutional neural network to obtain an environment feature vector; the hyperspectral information is input into the spectral information branch, and the spatial-spectral features are extracted using a three-dimensional convolutional neural network to obtain a spectral feature vector; the three-dimensional structure data is input into the three-dimensional structure branch, and the point cloud data is processed using a point cloud convolutional network to obtain a structure feature vector; The environmental feature vector, the spectral feature vector and the structural feature vector are fused by using a multi-head self-attention module to obtain a fused feature; the fused feature is input into a multi-layer perceptron, and a feature vector of a fixed dimension is output as a comprehensive digital model; Based on the comprehensive digital model, a clustering partitioning algorithm is used to intelligently partition the surface of the automobile parts to determine a plurality of sub-areas.

3. The method according to claim 2, characterized in that The cluster partitioning algorithm includes: Constructing a cluster partitioning algorithm based on K-means clustering; initializing cluster centers and using feature vectors in the comprehensive digital model as initial data points; Calculate the distance from each data point to each cluster center, introduce spatial constraints when calculating the distance, add spatial coordinate terms to the distance calculation formula, and assign adjustable weight coefficients; assign each data point to the nearest cluster center; calculate the average value of all data points in each cluster, determine the new cluster center, and update it; calculate the local curvature of each data point, and use the local curvature as the cluster feature dimension; Repeat the iteration until the position change of the cluster center is less than the preset cluster center change threshold, and the optimal partitioning scheme is obtained; Calculating the silhouette coefficient of the optimal partitioning scheme by a similarity measurement method; Within the preset range of partition numbers, the initialization and iterative calculation process is repeated to obtain the silhouette coefficients under different numbers of partitions, and the partition scheme with the largest silhouette coefficient is selected as the final intelligent partitioning result.

4. The method according to claim 1, characterized in that: Based on each sub-area and combined with the current environmental parameters, the intelligent agent trained by the reinforcement learning algorithm generates the optimal spraying parameters for each sub-area, including: Building a main network based on a deep Q network, the main network includes an input layer, two hidden layers and an output layer, wherein the number of nodes in the input layer is the same as the dimension of the state vector, the number of nodes in the output layer is equal to the size of the action space, the hidden layer uses a ReLU activation function, and the output layer uses a linear activation function; Based on the geometric features, material properties and environmental parameters of the sub-area, the state space is determined; based on the spraying parameters, the action space is determined; based on the preset coating quality indicators, the weighted sum of the coating quality indicators in the sub-area is calculated to construct a reward function; Based on the state space, action space and reward function, a simulation environment is constructed to simulate the spraying process of sub-areas on the surface of automobile parts; Construct an experience pool to store the experience data obtained by the interaction between the intelligent agent and the simulation environment, initialize the experience pool, create a target network based on the structure of the main network, and calculate the target Q value; The ε-greedy method is used to set the exploration strategy. The ε value is initially set to 0.9 and linearly decays to 0.1 as the number of training rounds increases. The intelligent agent selects spraying parameters according to the current sub-region state and exploration strategy, interacts with the simulation environment, obtains rewards and the next sub-region state, and calculates the temporal difference error based on the current sub-region state, the selected spraying parameters, the rewards obtained, the next sub-region state, and the target Q value calculated by the target network; The current sub-region state, the selected spraying parameters, the reward obtained, the next sub-region state and the time series difference error are taken as one piece of experience data, stored in and updated in the experience pool; a sampling probability is assigned to each piece of experience data based on the time series difference error in the experience pool, batch data is selected based on the sampling probability, and the loss function is calculated based on the batch data, the main network and the target network; Minimize the loss function through the gradient descent method, update the main network parameters, assign the weighted average of the main network parameters to the corresponding parameters of the target network, update the target network parameters, add Gaussian noise to the main network parameters and the target network parameters, and continue to iterate until the preset training rounds are reached. Determine the main network that has been trained, input the current state of each sub-area into the trained main network, and input the optimal spraying parameter combination of the sub-area.

5. The method according to claim 4, characterized in that Based on the optimal spraying parameters and the spatial relationship between sub-areas, the ant colony path planning algorithm is used to optimize the global spraying sequence and generate the optimal spraying path including: A node is generated for each sub-region. The edges between nodes are determined based on the spatial relationship between sub-regions. The graph structure is determined based on the nodes and edges. Based on the spatial distance between the sub-regions, the optimal spraying parameters corresponding to each sub-region and the surface curvature change of the sub-region, the weight of the edge is calculated, wherein the spatial distance is determined based on the Euclidean distance between adjacent sub-regions, the similarity between the corresponding vectors of the optimal spraying parameters of the two sub-regions is calculated by cosine similarity, the spraying parameter similarity is determined, the curvature between adjacent sub-regions is calculated, the curvature difference is determined, and the surface curvature change is obtained; According to the preset ant population size, the ant population and the corresponding pheromone matrix are initialized, and each ant randomly selects a sub-region as the starting sub-region; Based on the weight of the edge, determine the heuristic information, based on the pheromone matrix, determine the pheromone concentration, calculate the state transition probability according to the heuristic information and the pheromone concentration, determine the probability of each ant selecting the next sub-area, select the sub-area corresponding to the maximum probability, and move the ant to the corresponding sub-area; After the ant moves, the local pheromone of the edge passed by the ant during the movement is updated, and the local pheromone is updated based on the corresponding pheromone concentration in the pheromone matrix, the preset pheromone volatility coefficient and the preset initial pheromone concentration; Repeat the steps of sub-region selection and local pheromone update until all ants complete a complete path. Determine the path length of the ants based on the spatial distance between sub-regions and the time distance of the spray gun adjustment. Determine the corresponding spray quality score of the ants based on the similarity of spraying parameters and curvature changes between adjacent sub-regions. Update the global pheromone based on the path length and spray quality score. Repeat until the preset maximum number of iterations is reached, select the spray path corresponding to the maximum spray quality score, and determine the optimal spray path.

6. The method according to claim 5, characterized in that The path length is given by the following formula: ； Where L represents the total path length, m represents the number of sub-regions contained in the path, i represents the index of the sub-region in the path, r(·) represents the sub-region with the corresponding index in the path, d(·,·) represents the spatial distance between sub-regions, T represents the time coefficient required to adjust the spraying parameters, P(·) represents the optimal spraying parameter vector, and ||·|| represents the calculation of the Euclidean distance; The spraying quality score is as follows: ； Where Q represents, λ represents the weight coefficient, sim(·) represents the calculation of spraying parameter similarity, c(·) represents the calculation of average curvature, and |·| represents the absolute value of curvature difference.

7. The method according to claim 1, characterized in that A pre-built fuzzy logic controller is used to adjust the spraying parameters according to the deviation between the real-time quality evaluation result and the preset quality target until the spraying of the current automobile parts is completed, and the sprayed parts are obtained. The sprayed parts are scanned and tested to obtain the final spraying quality data. The final spraying quality data, environmental parameters, component characteristics and spraying parameters are integrated into a full-cycle comprehensive data set. Based on the full-cycle comprehensive data set, the gradient boosting decision tree algorithm is used to extract the spraying quality influencing variables. Based on the spraying quality influencing variables, the network parameters of the multimodal fusion neural network and the algorithm parameters of the reinforcement learning algorithm are updated, including: According to the deviation between the real-time quality assessment results and the preset quality targets, the spraying parameters are adjusted through the pre-built fuzzy logic controller until the current automotive parts spraying is completed, including: A dynamic fuzzy neural network is introduced, and the input and output variable membership function parameters of the fuzzy logic controller are adjusted according to the real-time spraying effect feedback, and the loss function is calculated, wherein the loss function includes the mean square error between the predicted quality and the actual quality, and the L2 regularization term of the difference before and after the membership function parameter adjustment; based on the loss function, the gradient is calculated by the back propagation algorithm, and the input and output variable membership function parameters of the fuzzy logic controller are updated in combination with the stochastic gradient descent method; the loss function calculation and parameter updating steps are repeated until the preset number of iterations is reached to obtain the sprayed parts; Use a high-precision three-dimensional laser scanner to perform an all-round scan on the spray-painted parts to generate a high-precision three-dimensional model; Based on the high-precision three-dimensional model, a multi-scale spraying quality assessment method is used to calculate a comprehensive quality score, the comprehensive quality score, environmental parameters, component characteristics and spraying parameters are integrated into a full-cycle comprehensive data set, and the full-cycle comprehensive data set is stored in a time series database; Based on the full-cycle comprehensive data set, the LightGBM algorithm is used to extract the variables affecting the spraying quality, and the SHAP value is used to explain the LightGBM model and quantify the importance of each feature to obtain the ranking of key factors affecting the spraying quality; Based on the spraying quality influencing variables, the network parameters of the multimodal fusion neural network are updated, wherein the multimodal fusion neural network includes a one-dimensional convolutional neural network for processing time series spraying parameter data, a two-dimensional convolutional neural network for processing component geometric feature images, a long short-term memory network for processing environmental parameter time series, and a multi-layer perceptron sub-network for processing static features, and the outputs of each sub-network are fused through a Transformer-based attention layer; Based on the spraying quality influencing variables, the algorithm parameters of the reinforcement learning algorithm are updated, and the updated multimodal fusion neural network and the reinforcement learning algorithm are applied to the spraying process of the next batch of automobile parts to continuously adjust the spraying quality.

8. An automobile parts spraying control system, used to implement the method according to any one of claims 1 to 7, characterized in that: include: The first unit is used to collect spraying environment parameters in real time through temperature and humidity sensors, air pressure sensors and air quality sensors; use multispectral cameras and three-dimensional structured light scanners to obtain surface spectral information and three-dimensional structural data of automobile parts; input spraying environment parameters, surface spectral information and three-dimensional structural data into a pre-trained multimodal fusion neural network to generate a comprehensive digital model containing environmental parameters and component characteristics, and based on the comprehensive digital model, use a clustering partitioning algorithm to intelligently partition the surface of automobile parts into multiple sub-areas; The second unit is used to generate the optimal spraying parameters for each sub-area based on each sub-area and in combination with the current environmental parameters, using an intelligent agent trained by a reinforcement learning algorithm, and based on the optimal spraying parameters and the spatial relationship between the sub-areas, using an ant colony path planning algorithm to optimize the global spraying sequence and generate the optimal spraying path; the optimal spraying parameters and the optimal spraying path are sent to a spraying robot with an adaptive control function, and during the spraying process, the spraying state data is collected in real time through a high-speed camera and a coating thickness sensor, and the spraying state data is input into a pre-trained convolutional neural network to generate a real-time quality evaluation result of the spraying; The third unit is used to use a pre-built fuzzy logic controller to adjust the spraying parameters according to the deviation between the real-time quality assessment result and the preset quality target until the spraying of the current automobile parts is completed, the sprayed parts are obtained, the sprayed parts are scanned and inspected, and the final spraying quality data is obtained. The final spraying quality data, environmental parameters, component characteristics and spraying parameters are integrated into a full-cycle comprehensive data set, based on the full-cycle comprehensive data set, the gradient boosting decision tree algorithm is used to extract the spraying quality influencing variables, and based on the spraying quality influencing variables, the network parameters of the multimodal fusion neural network and the algorithm parameters of the reinforcement learning algorithm are updated.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images