A method, system, medium, and product for autonomous optimization of a paint spraying process

Abstract

A method, system, medium and product for autonomous optimization of a paint spraying process, relating to the field of process optimization. The method comprises: inputting multi-modal sensing data in the paint spraying process into a pre-trained process diagnosis model to output a semantic diagnosis description of the current paint spraying process state; generating a control strategy scheme for defect elimination or tendency based on the semantic diagnosis description, workpiece information of a workpiece to be sprayed, and target paint film quality standards; inputting the control strategy scheme into a digital twin model for simulation, and based on the simulation prediction results, rolling optimization of a set of coordinated adjustment instructions in the control strategy scheme is performed through a dynamic optimizer to form a final optimized strategy. The implementation of the above technical solution realizes autonomous optimization of the paint spraying process.

Description

Technical Field

[0001] This application relates to the field of process optimization, and in particular to a method, system, medium, and product for autonomous optimization of spray painting processes. Background Technology

[0002] In the industrial manufacturing sector, painting is a crucial process, widely used in numerous industries such as automotive, aerospace, and machinery manufacturing. High-quality painting not only improves the appearance of products but also enhances their corrosion resistance and abrasion resistance, extending their service life. With the continuous development of the manufacturing industry, the demands for the quality and efficiency of painting processes are also increasing.

[0003] In existing spray painting processes, manual inspection is typically used to ensure paint quality. This involves observing whether environmental parameters meet requirements and identifying obvious paint film defects such as runs, orange peel, and bubbles. Once a problem is found, workers adjust the spray painting parameters based on their experience, such as adjusting the spray gun pressure, spray distance, and spray speed. However, manual inspection is highly subjective; different workers may have different judgment standards, leading to inconsistent paint quality. Furthermore, it can only monitor a limited number of environmental parameters, failing to comprehensively reflect the actual conditions during the spraying process and making it difficult to meet the complex and ever-changing demands of spray painting. When paint film defects occur, relying on worker experience for adjustments is often inaccurate and inefficient, failing to fundamentally solve the problem and potentially causing process fluctuations, further affecting paint quality. Summary of the Invention

[0004] This application provides a method, system, medium, and product for autonomous optimization of the spray painting process, which realizes autonomous optimization of the spray painting process, improves paint film quality, reduces energy consumption and material consumption, and continuously improves the accuracy and reliability of the system through model parameter adjustment.

[0005] Firstly, this application provides a method for autonomous optimization of a painting process, the method comprising:

[0006] Multimodal sensing data during the painting process is acquired in real time. The multimodal sensing data is input into a pre-trained process diagnostic model, and a semantic diagnostic description of the current painting process state is output. The multimodal sensing data includes spectral image data, acoustic emission data, and environmental field data. The semantic diagnostic description includes the identification of paint film defects or tendencies and the inference of physical causes.

[0007] Based on the semantic diagnostic description, the workpiece information of the workpiece to be coated, and the target paint film quality standard, a control strategy scheme for eliminating the defects or tendencies is generated. The control strategy scheme includes a set of coordinated adjustment instructions for multiple environmental control devices and coating actuators.

[0008] The control strategy scheme is input into a digital twin model for simulation to predict the evolution path of the coating state within a preset time after the execution of the control strategy scheme. Based on the simulation prediction results, the collaborative adjustment instruction set in the control strategy scheme is rolled optimization by a dynamic optimizer. The dynamic optimizer aims to minimize the weighted sum of quality deviation, energy consumption and material consumption, and takes into account the constraints of the dynamic response characteristics of each actuator to form the final optimization strategy.

[0009] The final optimization strategy is executed, and after execution, the strategy execution performance index is calculated based on the actual obtained final quality data of the coating film and the simulation prediction data. According to the strategy execution performance index, the diagnostic confidence parameters of the process diagnostic model and the dynamic response parameters of the digital twin model are backpropagated and adjusted.

[0010] By employing the above technical solution, multimodal sensor data is utilized to achieve real-time identification and physical cause inference of paint film defects and tendencies through a pre-trained model, elevating the process from phenomenon judgment to mechanism diagnosis. Combining diagnostic results, workpiece information, and quality standards, coordinated adjustment commands for various equipment and mechanisms are automatically generated, transforming control from single-point adjustment to systemic linkage optimization. Digital twins are used for strategy simulation and state evolution prediction, with rolling optimization guided by multiple objectives such as quality, energy consumption, and material consumption. The strategy is verified and dynamically improved before implementation, increasing the first-time success rate and overall efficiency of control. By comparing actual results with simulation predictions, performance indicators are calculated, and model parameters are adjusted in reverse, enabling the system to possess online learning and adaptive evolution capabilities, continuously improving diagnostic accuracy and control precision.

[0011] In some embodiments, the step of inputting the multimodal sensing data into a pre-trained process diagnostic model and outputting a semantic diagnostic description of the current painting process state specifically includes:

[0012] The spectral image data, the acoustic emission data, and the environmental field data are preprocessed, including timestamp alignment and spatial location registration.

[0013] The preprocessed spectral image data is subjected to pixel-by-pixel spectral analysis to extract the absorptivity and reflectivity distribution maps in the preset paint film characteristic bands to obtain visual spectral features. The preprocessed acoustic emission data is subjected to time-frequency analysis to extract the characteristic frequency band energy ratio characterizing the paint droplet size distribution and collision energy to obtain acoustic vibration features. The preprocessed environmental field data is spatially interpolated to generate a three-dimensional distribution cloud map of the temperature field, pressure field, and airflow velocity vector field in the spray booth to obtain environmental physical field features.

[0014] The visual spectral features, acoustic vibration features, and environmental physical field features are weighted and fused through the cross-modal attention fusion module to generate a deep working condition representation vector.

[0015] The semantic diagnostic description is output based on the deep working condition representation vector and the trained process knowledge graph.

[0016] By employing the aforementioned technical solution, specialized signal processing is performed on spectral, acoustic, and environmental data to extract essential features reflecting the visual characteristics of the paint film, spraying dynamics, and macroscopic environmental field. An attention mechanism is then used for adaptive weighted fusion, forming a comprehensive and accurate deep working condition representation vector, laying a data foundation for precise diagnosis. Combining this fused deep working condition representation vector with a process knowledge graph containing domain knowledge not only outputs defect identification results but also traces their physical causes, achieving a transformation from complex data to semantic description. This makes the diagnostic process and results highly interpretable and physically meaningful. Through a cross-modal attention fusion mechanism, the system can dynamically adjust the dependence weights on different data sources according to the current working conditions, enhancing the system's adaptability to complex and changing conditions and avoiding misjudgments caused by single sensor failure or environmental interference.

[0017] In some embodiments, outputting the semantic diagnostic description based on the deep working condition representation vector and the trained process knowledge graph includes:

[0018] Calculate the similarity between the deep working condition representation vector and the predefined feature vectors of all nodes in the process knowledge graph, and select multiple nodes with similarity higher than a first threshold as a candidate related node set. The nodes in the process knowledge graph are pre-labeled with node types, which include equipment status, environmental parameters, physical phenomena and defect types.

[0019] Taking the first node of the candidate related node set whose node type is defect type as the starting point of reasoning, in the process knowledge graph, a search with a limited step size is performed along the reverse of the causal relationship edge until the root cause node whose node type is equipment status or environmental parameter is found. All the searched nodes and connecting edges are combined in sequence into one or more directed paths from the root cause node to the first node, which are used as candidate causal reasoning paths.

[0020] For each of the candidate causal reasoning paths, the defect represented by the first node is determined as a specific defect type, and the location of the defect is determined based on the original spatial location information corresponding to the generated deep working condition representation vector.

[0021] Construct a physical cause chain by connecting all second nodes of type physical phenomenon along the candidate causal reasoning path from the root cause node to the first node in causal order.

[0022] Based on the average similarity between the deep working condition representation vector and the feature vectors of each node on the candidate causal reasoning path, and the prior weight of the candidate causal reasoning path in the knowledge graph, the logical confidence of the candidate causal reasoning path is calculated. Based on the logical confidence and the predefined defect severity of the first node, the severity level is calculated through a pre-defined mapping rule.

[0023] From all the candidate causal reasoning paths, the path with the highest logical confidence is selected, and the generated specific defect type, location of occurrence, physical cause chain and severity level are filled into the predefined natural language template to generate the semantic diagnostic description.

[0024] By employing the above technical solution, real-time data features are matched with a static process knowledge graph for similarity, and a reverse search is performed along the causal relationship edges in the graph. This allows for the systematic identification of the root cause (equipment status or environmental parameters) of defects, rather than focusing on surface phenomena, thus achieving traceability in diagnosis. The technical process automatically integrates information such as the identified defect type, specific location, complete physical causal chain, and quantitatively assessed severity level into a structured semantic description. The diagnostic process combines two paradigms: data-driven (based on similarity matching between deep representation vectors and nodes) and knowledge-driven (based on causal logical reasoning from a knowledge graph). The reliability of each reasoning path is quantified by calculating logical confidence, and data matching degree and prior knowledge weights are comprehensively considered to select the most probable causal path, enhancing the persuasiveness and anti-interference ability of the diagnostic results.

[0025] In some embodiments, generating a control strategy for eliminating the defects or tendencies based on the semantic diagnostic description, workpiece information of the workpiece to be coated, and target paint film quality standards includes:

[0026] The physical cause chain in the semantic diagnostic description is analyzed, and the starting node of the physical cause chain is identified as the cause node, and the node directly related to the defect or tendency is identified as the direct influencing factor node.

[0027] From the preset process countermeasure rule library, a set of corrective actions are matched for the cause node and the direct influencing factor node, respectively. The corrective action corresponds to the adjustment of a single parameter of a single environmental control device or spraying actuator.

[0028] Based on the geometric features and material properties in the workpiece information, and the weights of various indicators in the target paint film quality standard, all the matching correction actions are prioritized and conflict detected.

[0029] Based on the sorting and conflict detection results, a strategy skeleton containing multiple candidate atomic action sequences is generated, and the control strategy scheme is generated based on the strategy skeleton.

[0030] By employing the above technical solution, the root cause nodes and directly influencing factor nodes requiring regulation are accurately located through parsing the physical causal chain in semantic diagnosis. Corresponding corrective actions are then directly mapped from a pre-set rule base, ensuring a high degree of consistency and relevance between the control strategy and the diagnostic conclusions. Taking into account the workpiece's geometric features, material properties, and the weighting of quality standard indicators, the initially matched corrective actions are prioritized and conflict detected. This process ensures the feasibility of the generated control strategy under actual process constraints and enables it to adapt to the specific requirements of different workpieces.

[0031] In some embodiments, generating the control strategy scheme based on the strategy skeleton includes:

[0032] The corresponding device response model is invoked to predict the dynamic change curve of the target process parameters after each candidate atomic action sequence in the strategy skeleton is executed;

[0033] The dynamic change curve is compared with the ideal process parameter trajectory corresponding to the target coating quality standard, and the expected performance score of each candidate atomic action sequence is calculated.

[0034] The target candidate atomic action sequence with the highest expected performance score is selected as the basis, and each atomic action in the target candidate atomic action sequence is transformed into an execution instruction with execution time point, duration, target value and rate of change according to the rolling optimization requirements of the dynamic optimizer;

[0035] All the execution instructions are arranged in chronological order and embedded with inter-device synchronization and interlocking logic to form the control strategy scheme.

[0036] By employing the above technical solution, the execution effect of each candidate action sequence is dynamically simulated using an equipment response model, and quantitatively compared with the ideal process trajectory (calculating the expected performance score) to select the theoretically optimal implementation scheme. The selected optimal action sequence (atomic action) is concretized into executable instructions with precise time points, durations, target values, and rates of change. This transformation process endows the abstract control logic with precise time and quantitative dimensions, providing clear and unambiguous operating guidelines for automated equipment. During the instruction orchestration stage, synchronization and interlocking logic between devices is proactively embedded to ensure that multiple environmental control devices and spraying actuators can work collaboratively according to predetermined timing and logical relationships. This avoids process disruptions or equipment safety issues caused by action conflicts or timing errors, improving the reliability of collaborative control of complex systems.

[0037] In some embodiments, calculating the strategy execution performance index based on the actually acquired final paint film quality data and simulation prediction data specifically includes:

[0038] The same set of quality features are extracted from the final quality data of the paint film and the simulation prediction data, respectively. The quality features include average film thickness, film thickness uniformity, gloss, color difference and the area ratio of defect areas.

[0039] For each quality feature, the deviation between the actual value and the simulation prediction value is calculated, and the deviation is normalized to the corresponding tolerance range to obtain the individual deviation degree of the quality feature.

[0040] Based on the weighting coefficients of each quality characteristic preset in the target paint film quality standard, the individual deviations are weighted and summed to obtain the overall quality conformity index.

[0041] By combining the energy and material cost data consumed in executing the final optimization strategy, the overall quality conformity index and unit cost are comprehensively calculated to generate the strategy execution performance index.

[0042] By employing the aforementioned technical solution, key quality characteristics such as average film thickness, uniformity, gloss, color difference, and defect area are extracted. The measured values ​​are then compared with the simulated predicted values ​​item by item to achieve comprehensive and quantitative benchmarking of paint film quality, from macroscopic thickness to microscopic appearance, and from uniformity to defects. This multi-dimensional evaluation provides a more comprehensive reflection of the strategy's actual effectiveness than a single indicator. The calculated strategy execution performance index is a quantified, multi-dimensional feedback condensed signal. It accurately characterizes the predictive accuracy of the digital twin model (reflected in quality fit) and the actual economic efficiency of the strategy generated by the dynamic optimizer. This index will become the core basis for subsequent parameter back-adjustment of the diagnostic model and the twin model, driving continuous learning and improvement of the entire system.

[0043] In some embodiments, the backpropagation and adjustment of the diagnostic confidence parameters of the process diagnostic model and the dynamic response parameters of the digital twin model includes:

[0044] Based on the strategy execution performance index, the overall loss value of model parameter adjustment is calculated. The overall loss value is composed of a weighted average of the quality prediction error term, the energy consumption prediction error term, and the diagnostic accuracy error term.

[0045] The gradient of the dynamic response parameters of each actuator and environmental control device in the digital twin model is calculated based on the overall loss value, and the dynamic response parameters are updated using an adaptive learning rate algorithm according to the gradient direction and magnitude.

[0046] The control strategy scheme is re-simulated based on the updated digital twin model to generate corrected simulation data, and the confidence deviation of each inference node of the process diagnostic model in the semantic diagnostic description is calculated based on the simulation data.

[0047] Based on the confidence bias, the feature weights of the cross-modal attention fusion module in the process diagnostic model and the prior weights of the causal relationships related to the defects or tendencies in the process knowledge graph are adjusted by the backpropagation algorithm.

[0048] By adopting the above technical solution, the system transforms comprehensive performance indicators, including quality, energy consumption, and diagnostic accuracy, into overall loss values. This enables unified and guided joint optimization of the parameters of the upstream digital twin model and process diagnostic model, starting from the final execution result. This breaks through the limitations of traditional independent adjustment of single modules, ensuring that each model evolves collaboratively towards improving the overall system performance. By updating the dynamic response parameters of each device (actuator and environmental control equipment) in the digital twin model based on gradients, the deviation between model predictions (such as process parameter change curves) and the actual physical response of the equipment can be continuously corrected. This allows the digital twin model to track factors such as equipment aging and environmental changes, maintaining the long-term accuracy and reliability of its simulation predictions, which is crucial for the continued value of digital twins. The updated, more accurate digital twin model is used to re-evaluate the diagnostic results and calculate the confidence bias of the diagnostic model at each inference node. By adjusting the weights of the cross-modal attention fusion module through backpropagation, the model can more accurately rely on key data sources. Simultaneously, adjusting the prior weights of relevant causal relationships in the process knowledge graph essentially involves dynamically correcting and enhancing the domain knowledge base based on practical feedback. This enables the diagnostic model to achieve self-evolution at both the data and knowledge levels.

[0049] In a second aspect, embodiments of this application provide a computer system including a memory, a processor, and a computer program stored in the memory; the processor executes the computer program to implement the steps of the method described in any possible implementation of the first aspect.

[0050] Thirdly, embodiments of this application provide a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implement the steps of the method described in any possible implementation of the first aspect.

[0051] Fourthly, embodiments of this application provide a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the method described in any possible implementation of the first aspect.

[0052] It is understood that the computer system provided in the second aspect, the storage medium provided in the third aspect, and the computer program product provided in the fourth aspect are all used to execute the method provided in this application. Therefore, the beneficial effects they can achieve can be referred to the beneficial effects in the corresponding methods, and will not be repeated here.

[0053] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages:

[0054] 1. By integrating multimodal data such as spectral, acoustic emission, and environmental field, and using process diagnostic models for semantic analysis, not only can paint film defects be identified, but their physical causes can also be traced, providing a clear basis for precise control.

[0055] 2. Before the actual execution of the control strategy, a digital twin model is used for simulation to predict the evolution path of the paint film state. This allows the system to preview the strategy's effects in a virtual space, identify potential problems in advance, and thus avoid quality risks in actual production.

[0056] 3. The dynamic optimizer performs rolling optimization with the comprehensive goal of minimizing quality deviation, energy consumption, and material consumption, ensuring that the final strategy is the most cost-effective solution while meeting quality requirements, thus achieving a balance between quality and efficiency.

[0057] 4. By comparing the actual paint film quality data after the strategy is implemented with the simulation prediction data, the performance index is calculated, and the key parameters of the process diagnostic model and the digital twin model are backpropagated and adjusted accordingly. This mechanism enables the diagnostic model to continuously improve its diagnostic accuracy and confidence, and the digital twin model to continuously calibrate its dynamic response characteristics to better fit the actual physical equipment. Attached Figure Description

[0058] Figure 1This is a flowchart illustrating a method for autonomous optimization of a painting process in an embodiment of this application.

[0059] Figure 2 This is a flowchart illustrating the output of a semantic diagnostic description of the current painting process status in an embodiment of this application.

[0060] Figure 3 This is a schematic diagram of an exemplary hardware structure of a computer system in an embodiment of this application. Detailed Implementation

[0061] The terminology used in the following embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification and appended claims of this application, the singular expressions “a,” “an,” “the,” “the,” “the,” and “this” are intended to include the plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this application refers to any or all possible combinations including one or more of the listed items.

[0062] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature, and in the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.

[0063] The following is combined Figure 1 The method of the embodiments of this application will be described below.

[0064] Figure 1 This is a flowchart illustrating a method for autonomous optimization of a painting process according to an embodiment of this application, as shown below. Figure 1 As shown, the method includes the following steps:

[0065] S101. Acquire multimodal sensing data in real time during the painting process, input the multimodal sensing data into a pre-trained process diagnostic model, and output a semantic diagnostic description of the current painting process state. The multimodal sensing data includes spectral image data, acoustic emission data, and environmental field data. The semantic diagnostic description includes the identification of paint film defects or tendencies and the inference of physical causes.

[0066] S102. Based on the semantic diagnostic description, the workpiece information of the workpiece to be sprayed, and the target paint film quality standard, a control strategy scheme for eliminating the defects or tendencies is generated. The control strategy scheme includes a set of coordinated adjustment instructions for multiple environmental conditioning devices and spraying actuators.

[0067] S103. Input the control strategy scheme into the digital twin model for simulation, predict the evolution path of the paint film state within a preset time after the execution of the control strategy scheme, and based on the simulation prediction results, perform rolling optimization of the collaborative adjustment instruction set in the control strategy scheme through a dynamic optimizer. The dynamic optimizer aims to minimize the weighted sum of quality deviation, energy consumption and material consumption, and takes into account the constraints of the dynamic response characteristics of each actuator to form the final optimization strategy.

[0068] S104. Execute the final optimization strategy, and after execution, calculate the strategy execution performance index based on the actual obtained final quality data of the coating film and the simulation prediction data. According to the strategy execution performance index, backpropagate and adjust the diagnostic confidence parameters of the process diagnostic model and the dynamic response parameters of the digital twin model.

[0069] The system synchronously collects multimodal sensing data during the painting process using various types of sensors. This multimodal sensing data includes spectral image data, acoustic emission data, and environmental field data. Spectral image data captures the optical properties of the paint film, reflecting film thickness, uniformity, color, gloss, and early microscopic defects (such as orange peel and pinholes). Acoustic emission data monitors mechanical vibrations and sound signals during the spraying process, indirectly reflecting droplet size distribution, atomization state, nozzle condition, and microscopic events during coating adhesion. Environmental field data includes temperature, humidity, wind speed, and wind pressure within the spray booth; these data collectively constitute the key environmental physical fields affecting paint film drying, leveling, and final quality. This multimodal sensing data is then input into a pre-trained process diagnostic model. This model is an AI (artificial intelligence) model that integrates deep learning and knowledge reasoning. First, it extracts high-level features from various data sources (such as texture in images, characteristic frequencies in sound spectra, and gradients of the environmental field). Then, through cross-modal learning mechanisms (such as attention networks), it fuses these features into a unified deep state vector that comprehensively represents the current working condition. Based on this state vector, the model can not only identify specific paint film defects (such as sagging, particles, and whitening) or defect tendencies, but more importantly, it can infer the physical causes of these problems. For example, the diagnosis is not simply insufficient gloss, but rather, "Due to the low temperature in the spray booth (root cause), the solvent evaporates too slowly, causing the paint film leveling time to be too long (phenomenon), ultimately resulting in orange peel defects (result)." This output is a semantic diagnostic description, which connects the phenomenon and the mechanism, providing a clear direction for subsequent precise control.

[0070] Workpiece information includes its geometry, dimensions, and material (e.g., metal, plastic), which determines the spraying trajectory, paint adhesion characteristics, and heat capacity. Target paint film quality standards define quantitative indicators and their weights for quality requirements (e.g., film thickness tolerance ±5μm, gloss >90GU, with uniformity having the highest weight). Based on semantic diagnostic descriptions, workpiece information, and target paint film quality standards, the system generates a preliminary control strategy from a pre-defined process countermeasure rule base or through optimization algorithms. This strategy is not a single instruction but a coordinated adjustment instruction set, simultaneously planning the actions of multiple execution units. For example, environmental control equipment might increase heater power (raise temperature), decrease humidifier output (reduce humidity), and increase exhaust frequency (adjust air pressure); the spraying execution mechanism might adjust robot spraying speed, change spray gun atomization pressure, and fine-tune paint flow rate. The generated strategy must ensure that these adjustments are coordinated temporally and logically to avoid conflicts.

[0071] The generated control strategy is input into a digital twin model of the painting process. The digital twin model is a high-fidelity virtual mapping that includes a spray booth physics model, a paint dynamics model, and a robot motion model. The digital twin model simulates the execution of the control strategy and predicts the evolution path of the paint film over a future period (the next painting cycle), such as how the film thickness increases and how the surface morphology changes. The dynamic optimizer receives the simulation prediction results and compares them to the ideal quality target trajectory. The optimizer's objective function is to minimize the weighted sum of quality deviation, energy consumption, and material consumption, meaning it seeks the optimal balance between ensuring quality and saving costs. The optimizer considers the dynamic response characteristics of the actuators (such as the time required for fan acceleration and the thermal inertia of the heater), using achievable adjustment rates and magnitudes as constraints, iteratively adjusting and resimulating the instruction set. This process may repeat multiple times, like "rolling," ultimately outputting a final optimized strategy that has been validated in the virtual simulation and balances quality and efficiency.

[0072] The final optimized strategy is then distributed to the environmental control equipment and spraying actuators on the actual production line, driving them to coordinate their actions according to the plan. After the strategy is executed, the final quality data of the paint film is obtained through detection equipment (such as film thickness gauges, gloss meters, and vision inspection systems). This actual data is compared with the simulation prediction data from the digital twin to calculate the strategy execution efficiency index. This index comprehensively reflects the actual effect of the strategy in improving quality, saving energy and materials. Based on the calculated efficiency index, the system automatically performs backpropagation and adjustment on the key parameters of the two core models. Process diagnostic model: Adjust its diagnostic confidence parameters. If the diagnosis is accurate but the effect after adjustment is not good, the problem inference may be biased; if the diagnosis is wrong, the weight of relevant features needs to be reduced. This makes the diagnostic model more and more accurate. Digital twin model: Adjust its dynamic response parameters. If the simulation prediction deviates significantly from the actual situation, it indicates that the model's characterization of the equipment or physical process is inaccurate, and its response curve, time constant, and other parameters need to be calibrated. This makes the digital twin model closer to reality.

[0073] Figure 2 This is a flowchart illustrating the output of a semantic diagnostic description of the current painting process status in an embodiment of this application, such as... Figure 2 As shown, the step of inputting the multimodal sensing data into a pre-trained process diagnostic model and outputting a semantic diagnostic description of the current painting process status specifically includes:

[0074] S201. Preprocess the spectral image data, the acoustic emission data, and the environmental field data, wherein the preprocessing includes timestamp alignment and spatial location registration;

[0075] S202. Perform pixel-by-pixel spectral analysis on the preprocessed spectral image data to extract the absorptivity and reflectivity distribution maps in the preset paint film characteristic bands to obtain visual spectral features. Perform time-frequency analysis on the preprocessed acoustic emission data to extract the characteristic frequency band energy ratio that characterizes the paint droplet size distribution and collision energy to obtain acoustic vibration features. Perform spatial interpolation on the preprocessed environmental field data to generate a three-dimensional distribution cloud map of the temperature field, pressure field, and airflow velocity vector field inside the spray booth to obtain environmental physical field features.

[0076] S203. Through the cross-modal attention fusion module, the visual spectral features, the acoustic vibration features and the environmental physical field features are weighted and fused to generate a deep working condition representation vector.

[0077] S204. Output the semantic diagnostic description based on the deep working condition representation vector and the trained process knowledge graph.

[0078] Spectroscopic cameras, acoustic emission sensors, and temperature / humidity / wind speed sensors may sample at different frequencies (e.g., 30Hz, 10kHz, 1Hz). The system establishes a unified time reference (e.g., a high-precision system clock) to timestamp all data streams synchronously. Before feature extraction, interpolation or resampling techniques are used to align all data to a unified time series, ensuring that the analysis focuses on the same working conditions at the same moment. Different sensors monitor different spatial points. For example, a spectral image corresponds to a certain area on the workpiece surface, an acoustic emission sensor is installed on a spray gun or at a specific location, and environmental sensors are distributed at different points within the booth. A unified three-dimensional spatial coordinate system for the paint booth is established. The measurements of each sensor are correlated with their monitored physical spatial location. For example, each pixel of the spectral image is mapped to the three-dimensional coordinates of the workpiece surface; the discrete point measurements of the environmental sensors are then interpolated to form a continuous field distribution. This ensures that the subsequently extracted features are comparable and correlated in a physical spatial sense. Visual spectral feature extraction: Pixel-by-pixel spectral analysis is performed on the registered spectral image. For key wavelengths preset based on paint characteristics (e.g., near-infrared bands are sensitive to film thickness, and specific visible light bands are sensitive to color), the absorptivity and reflectivity of each pixel in that band are calculated. Absorptivity distribution maps (reflecting film thickness and material uniformity) and reflectivity distribution maps (reflecting color, gloss, and surface texture) are generated. This directly quantifies the visual appearance and microstructure of the paint film from an optical perspective, providing direct evidence for detecting defects such as orange peel, runs, and color differences. Acoustic vibration feature extraction: Time-frequency analysis (e.g., short-time Fourier transform, wavelet transform) is performed on acoustic emission data (usually time-domain waveforms) to convert it into a time-spectrum that simultaneously reflects the changes in frequency components and energy over time. In the time-spectrum, the energy ratio of characteristic frequency bands related to the physical mechanism of the spraying process is identified and extracted. Droplet size distribution: When droplets of different sizes impact the workpiece surface, they generate sound waves within a specific frequency range. Changes in the energy ratio of characteristic frequency bands can indirectly reflect the atomization quality. Collision energy: The total energy of the spraying impact is reflected in the total energy of the acoustic emission or specific mid-to-high frequency components. These acoustic vibration characteristics are unique and sensitive indicators for monitoring the spraying dynamics process (atomization state, nozzle clogging, gun distance changes). Environmental physical field feature extraction: Spatial interpolation (e.g., Kriging interpolation, radial basis function interpolation) is performed on data from discretely arranged environmental sensors (temperature, pressure, anemometers). This generates a three-dimensional distribution cloud map of the temperature field, pressure field, and airflow velocity vector field within the spray booth. This is no longer a few isolated readings, but a complete and visual description of the macroscopic physical state of the entire working environment. It can show whether there are temperature dead zones, airflow vortices, or pressure imbalances, which are macroscopic environmental root causes leading to uneven paint film drying, particle formation, or pinholes. The cross-modal attention fusion module is a neural network structure. It receives the above three types of features as input.The module automatically learns and calculates the importance weights (attention weights) of visual, acoustic, and environmental field features for the final diagnosis under the current working condition. For example, if the current problem is suspected to be "orange peel," the module may pay more attention to visual spectral features (surface texture) and environmental physical field features (temperature, airflow); if it is suspected to be "poor atomization," it may pay more attention to acoustic vibration features and environmental physical field features (air pressure). Based on these dynamic weights, the module weights and deeply interacts with the three features to generate a single, high-dimensional deep working condition representation vector. This vector condenses the essential information of the current spraying state in multi-dimensional space. Process Knowledge Graph: This is a pre-built, structured process knowledge base. Its nodes represent various entities or states (such as "low spray gun atomization pressure," "low room temperature," "orange peel paint film," "slow solvent evaporation"), and edges represent the causal relationships between them (such as "low pressure" → "poor atomization" → "large paint droplet size" → "orange peel"). The system matches or calculates the similarity between the generated deep working condition representation vector and the feature patterns corresponding to the nodes in the knowledge graph. By traversing and reasoning within the graph (e.g., searching backwards along causal edges from the matched "defect phenomenon" node to find possible "root cause" nodes), a semantic diagnostic description is ultimately generated. Example output: "An orange peel defect was identified in area A of the workpiece (identification). The cause is inferred to be insufficient atomization pressure of spray gun X, resulting in excessively large paint droplet size (physical cause 1), coupled with excessively high airflow velocity above this area, accelerating solvent evaporation (physical cause 2), resulting in the combined effect."

[0079] In some embodiments, outputting the semantic diagnostic description based on the deep working condition representation vector and the trained process knowledge graph includes:

[0080] Calculate the similarity between the deep working condition representation vector and the predefined feature vectors of all nodes in the process knowledge graph, and select multiple nodes with similarity higher than a first threshold as a candidate related node set. The nodes in the process knowledge graph are pre-labeled with node types, which include equipment status, environmental parameters, physical phenomena and defect types.

[0081] Taking the first node of the candidate related node set whose node type is defect type as the starting point of reasoning, in the process knowledge graph, a search with a limited step size is performed along the reverse of the causal relationship edge until the root cause node whose node type is equipment status or environmental parameter is found. All the searched nodes and connecting edges are combined in sequence into one or more directed paths from the root cause node to the first node, which are used as candidate causal reasoning paths.

[0082] For each of the candidate causal reasoning paths, the defect represented by the first node is determined as a specific defect type, and the location of the defect is determined based on the original spatial location information corresponding to the generated deep working condition representation vector.

[0083] Construct a physical cause chain by connecting all second nodes of type physical phenomenon along the candidate causal reasoning path from the root cause node to the first node in causal order.

[0084] Based on the average similarity between the deep working condition representation vector and the feature vectors of each node on the candidate causal reasoning path, and the prior weight of the candidate causal reasoning path in the knowledge graph, the logical confidence of the candidate causal reasoning path is calculated. Based on the logical confidence and the predefined defect severity of the first node, the severity level is calculated through a pre-defined mapping rule.

[0085] From all the candidate causal reasoning paths, the path with the highest logical confidence is selected, and the generated specific defect type, location of occurrence, physical cause chain and severity level are filled into the predefined natural language template to generate the semantic diagnostic description.

[0086] Node Feature Vectors: Each node in the process knowledge graph (e.g., "spray gun clogging," "low temperature," "coating pinholes") has a predefined feature vector. This vector encodes the typical pattern of the state or phenomenon represented by the node in the multi-dimensional feature space. The system calculates the similarity (e.g., cosine similarity) between the real-time generated deep condition representation vector and the feature vectors of all nodes in the knowledge graph. Multiple nodes with similarity higher than a preset first threshold are selected to form a candidate relevant node set. This set contains all potentially relevant process states, phenomena, and defects under the current data state. Node types (equipment state, environmental parameters, physical phenomena, defect types) provide semantic classification for subsequent reasoning. In the candidate relevant node set, nodes with the node type "defect type" (called the first node) are determined as the starting point for reasoning. For example, "orange peel," "sagging," and "particles." Starting from these first nodes, the search proceeds in the process knowledge graph along the reverse direction of the causal relationship edges. The edges of the knowledge graph typically point to "cause → effect," and the reverse search seeks the cause of this defect. Limit Step Size: To avoid excessively long search paths or loops, a maximum number of search steps is set. Termination Condition: The search continues until a node of type "Equipment Status" or "Environmental Parameter" is found. These two types of nodes represent directly controllable root causes (e.g., "abnormal robot spraying speed," "excessive humidity in the spray booth"). All nodes and edges traversed during the search process are combined in order from the root cause node (equipment / environment) to the defect node (first node) to form one or more directed causal paths. Each path represents a complete logical explanation from an operable root cause to the final defect. Defect Type: The defect represented by the first node (defect type node) at the end of the path is directly taken. Defect Location: Tracing back to the original data source that generated the deep working condition characterization vector (e.g., pixel coordinates in a spectral image, installation location of an acoustic emission sensor), the specific physical location of the defect in the actual workpiece or spray booth is determined through spatial mapping (e.g., "right side of the workpiece front end," "lower left corner of the spray booth"). On each causal path, all nodes of type "physical phenomenon" (called second nodes) are extracted, such as "solvent evaporation too fast," "paint droplet size too large," and "insufficient paint film leveling time." These nodes are then linked together according to their causal relationship from root cause to effect in the path, forming a logically coherent physical causal chain that describes the physical mechanism of the defect. A logical confidence score is calculated for each candidate causal reasoning path. This value is determined by two combined factors: data support: calculating the average similarity between the deep working condition representation vector and the feature vectors of all nodes on the path; the higher the average similarity, the more the current data supports each state in this causal chain; and prior knowledge weight: each causal edge or path in the knowledge graph can have a prior weight, assigned by historical data or expert experience, representing the prevalence or strength of the causal relationship under normal circumstances.Logical confidence is a function of data support and prior knowledge weights (e.g., a weighted average), quantifying "to what extent the current data supports this causal explanation." Each "defect type" node has a predefined base severity score (e.g., "orange peel" is 5 points, "color difference" is 3 points). Based on the calculated logical confidence of the path and the base severity of the defect, the system calculates the final severity level (e.g., high, medium, low) using a predefined mapping rule (e.g., a formula or lookup table). Severe defects with higher confidence are rated higher. From all candidate causal inference paths, the path with the highest logical confidence is selected as the most likely explanation. The system predefines a series of natural language templates, such as: "[Defect type] was detected at [location], with a severity level of [level]. The inferred cause is: [root cause node] leads to [physical phenomenon node 1], which in turn causes [physical phenomenon node 2], ultimately forming this defect." The specific defect type, location, physical cause chain, and severity level extracted from the selected path are filled into the corresponding positions in the template to generate a complete natural language description that conforms to human reading habits. For example: "An orange peel defect was detected in the middle of the surface of workpiece A, with a severity level of high. The reason is inferred to be: due to the low temperature of the spray booth, the solvent evaporation rate is too slow, which in turn causes the paint film leveling time to be abnormally prolonged, ultimately forming an orange peel texture."

[0087] In some embodiments, generating a control strategy for eliminating the defects or tendencies based on the semantic diagnostic description, workpiece information of the workpiece to be coated, and target paint film quality standards includes:

[0088] The physical cause chain in the semantic diagnostic description is analyzed, and the starting node of the physical cause chain is identified as the cause node, and the node directly related to the defect or tendency is identified as the direct influencing factor node.

[0089] From the preset process countermeasure rule library, a set of corrective actions are matched for the cause node and the direct influencing factor node, respectively. The corrective action corresponds to the adjustment of a single parameter of a single environmental control device or spraying actuator.

[0090] Based on the geometric features and material properties in the workpiece information, and the weights of various indicators in the target paint film quality standard, all the matching correction actions are prioritized and conflict detected.

[0091] Based on the sorting and conflict detection results, a strategy skeleton containing multiple candidate atomic action sequences is generated, and the control strategy scheme is generated based on the strategy skeleton.

[0092] The semantic diagnostic description received by the system contains a structured physical causal chain (e.g., low spray booth temperature (A) → slow solvent evaporation rate (B) → abnormally prolonged paint film leveling time (C) → orange peel defect (D)). The system identifies the starting node of the physical causal chain, i.e., the root cause of the chain. This is typically an equipment status or environmental parameter node (e.g., low spray booth temperature (A)). It also identifies nodes directly related to the final defect or tendency, which are usually physical phenomenon nodes adjacent to the defect node (e.g., abnormally prolonged paint film leveling time (C)). The system uses a process countermeasure rule base: this is a knowledge base pre-built by process experts and control system engineers. Its core is the mapping rule "IF <status / parameter node> THEN <corrective action set>". For example, IF node = low spray booth temperature THEN action set = {increase heater A power setting by 10%, turn on auxiliary heating unit B}. Based on the identified cause node and direct influencing factor node, the system retrieves and matches a corresponding set of corrective actions from the rule base. Each corrective action is "atomic," meaning it corresponds to an adjustment command for a single parameter (such as power, frequency, speed, or pressure) of a single environmental control device (e.g., heater, fan, humidifier) ​​or spraying actuator (e.g., spray gun, robot). This ensures the simplicity, clarity, and feasibility of the actions. Prioritization is based on: 1. Workpiece Information: The necessity or sensitivity of certain actions is determined based on the geometric characteristics (e.g., complex curved surfaces, thin-walled parts) and material properties (e.g., plastics, aluminum alloys) of the workpiece to be sprayed. For example, for heat-sensitive plastic workpieces, actions that increase temperature should have a lower priority or limited scope. 2. Quality Standard Weighting: Based on the weights of various indicators in the target paint film quality standard (e.g., gloss weight 0.4, film thickness uniformity weight 0.3), actions that significantly improve high-weight indicators are prioritized. Logical or physical conflicts between different actions are checked. Resource conflicts: Two actions require the same equipment to perform mutually exclusive operations (e.g., simultaneously increasing and decreasing fan speed). Effect Conflict: Two actions may have opposite effects on the same quality indicator (e.g., one action increases temperature to promote leveling, while the other decreases temperature to reduce bubbles caused by excessive solvent evaporation). The system needs to detect these conflicts and arbitrate them based on priority and process principles (e.g., discarding one of the conflicting actions or modifying its parameters). Based on the sorting and conflict detection results, the system combines the retained, optimized corrective actions into multiple candidate atomic action sequences; this sequence is the strategy skeleton. The skeleton may contain parallel branches (simultaneously adjusting multiple independent devices) and serial steps (adjusting A first, then B). Based on the strategy skeleton, the system generates a preliminary control strategy scheme. This scheme is a structured instruction set framework that clarifies: the list of all devices to be mobilized, the direction and approximate magnitude of parameter adjustments for each device, and the rough timing or logical relationship between actions (e.g., environmental adjustment actions take precedence over spraying actions).

[0093] In some embodiments, generating the control strategy scheme based on the strategy skeleton includes:

[0094] The corresponding device response model is invoked to predict the dynamic change curve of the target process parameters after each candidate atomic action sequence in the strategy skeleton is executed;

[0095] The dynamic change curve is compared with the ideal process parameter trajectory corresponding to the target coating quality standard, and the expected performance score of each candidate atomic action sequence is calculated.

[0096] The target candidate atomic action sequence with the highest expected performance score is selected as the basis, and each atomic action in the target candidate atomic action sequence is transformed into an execution instruction with execution time point, duration, target value and rate of change according to the rolling optimization requirements of the dynamic optimizer;

[0097] All the execution instructions are arranged in chronological order and embedded with inter-device synchronization and interlocking logic to form the control strategy scheme.

[0098] The system establishes dynamic response models for each key environmental control device (such as heaters, fans, and humidifiers) and spraying actuator (such as servo motors and spray gun solenoid valves). These models describe the dynamic relationship between the input commands (such as setpoints) and output physical quantities (such as actual temperature, wind speed, and spray flow rate) of the device, typically considering characteristics such as inertia, delay, and nonlinearity (e.g., heater heating is a first-order inertial process). Each candidate atomic action sequence in the strategy skeleton (e.g., sequence A = "second 1: heater power +10%; second 3: fan frequency +5%)" is used as input to drive these device response models for joint simulation. The simulation outputs a series of dynamic change curves for target process parameters. These parameters are mediating variables that directly affect paint film quality, such as the temperature curve at the target point in the spray booth, the wind speed curve on the workpiece surface, and the instantaneous paint flow rate curve. Based on the target paint film quality standards, a corresponding ideal process parameter trajectory can be derived or learned. For example, to achieve optimal leveling, the workpiece surface temperature should follow a curve that first rapidly rises to T1 and then slowly remains at T2 within a spraying cycle. This trajectory is a process-oriented and parameterized representation of the quality objective. For each candidate action sequence, the simulated dynamic change curve is compared point-by-point with the corresponding ideal trajectory. Typically, the integral of the deviation between the two (e.g., root mean square error RMSE) is calculated. The smaller the deviation, the stronger the ability of the action sequence to guide the process parameters to the ideal state, and the higher its expected performance score. The scoring formula may comprehensively consider multiple dynamic performance indicators such as tracking accuracy, settling time, and overshoot. The atomic action sequence with the highest expected performance score is selected as the target candidate atomic action sequence. Based on the rolling optimization requirements of the dynamic optimizer (meaning these requirements can be adjusted), each atomic action in the sequence is transformed from a rough description into a precise executable instruction. Each instruction must explicitly include the execution time point, duration, target value, and rate of change. All generated execution instructions are precisely arranged on a unified time axis according to their execution time point and duration, forming a time-synchronized instruction scheduling table. Synchronization logic: ensures that dependent actions are executed sequentially. For example, "start the spraying robot only after the spray booth temperature stabilizes at the set value." Interlocking logic: prevents conflicting actions from occurring simultaneously. For example, during the execution of the "clean spray gun" command, a "do not spray" logic lock is automatically inserted. The arranged command scheduling table and embedded logic are encapsulated into a structured, independently transmittable and parsable data object, which is the final control strategy scheme. This scheme can be directly issued to the workshop's scheduling and execution system or the controllers of individual devices.

[0099] In some embodiments, calculating the strategy execution performance index based on the actually acquired final paint film quality data and simulation prediction data specifically includes:

[0100] The same set of quality features are extracted from the final quality data of the paint film and the simulation prediction data, respectively. The quality features include average film thickness, film thickness uniformity, gloss, color difference and the area ratio of defect areas.

[0101] For each quality feature, the deviation between the actual value and the simulation prediction value is calculated, and the deviation is normalized to the corresponding tolerance range to obtain the individual deviation degree of the quality feature.

[0102] Based on the weighting coefficients of each quality characteristic preset in the target paint film quality standard, the individual deviations are weighted and summed to obtain the overall quality conformity index.

[0103] By combining the energy and material cost data consumed in executing the final optimization strategy, the overall quality conformity index and unit cost are comprehensively calculated to generate the strategy execution performance index.

[0104] The system pre-defines a set of core, quantifiable quality characteristics that collectively define several key aspects of paint film quality. These include: average film thickness (reflecting the overall thickness of the coating coverage, directly impacting protective performance and cost); film thickness uniformity (e.g., standard deviation, range) (reflecting the uniformity of thickness distribution, a core indicator for measuring the stability of the spraying process); gloss (the optical properties of the paint film surface, directly affecting appearance quality); color difference (measuring the difference between the actual color and a standard color sample); and the area ratio of defective regions (quantifying the severity of macroscopic or microscopic defects such as particles, pinholes, and runs). After the strategy is executed, the actual values ​​of these characteristics are extracted from the final paint film quality data obtained from actual testing equipment (e.g., online film thickness gauges, gloss meters, machine vision systems). Predicted values ​​for the same set of characteristics are extracted from the simulation prediction data output by the digital twin simulation system. For each quality characteristic, the absolute or relative deviation between the actual value and the simulation prediction value is calculated. For example, if the predicted average film thickness is 85 μm and the actual value is 82 μm, the absolute deviation is 3 μm. The calculated raw deviations are normalized to the tolerance range corresponding to the quality characteristic. The tolerance range, derived from process specifications or quality standards, defines the upper and lower limits of acceptance for that characteristic. For example, if the film thickness tolerance is 80μm ± 5μm, then the tolerance range is 10μm. The normalization formula, a common approach, is: Individual Deviation = |Actual Value - Predicted Value| / Tolerance Range. Pre-defined weighting coefficients for each quality characteristic are read from the target paint film quality standard. This reflects the importance the customer or process places on each indicator (e.g., for appearance parts, gloss and color difference have high weights; for corrosion-resistant parts, film thickness and uniformity have high weights). The overall quality compliance index is calculated using a weighted summation formula. During strategy execution, the system simultaneously records accurate data on consumed energy (e.g., electricity, compressed air) and materials (e.g., paint, solvents). The total cost is typically converted into the strategy execution cost per unit product or the strategy execution cost per unit sprayed area. The overall quality fit index (reflecting "effectiveness") and unit cost (reflecting "efficiency") are combined using a predefined function. A typical calculation method is: Strategy Execution Effectiveness Index = Overall Quality Fit Index + λ × Unit Cost. Here, λ is a trade-off coefficient used to adjust the relative importance of quality and cost. The smaller the value of this index, the higher the overall effectiveness (i.e., accurate quality prediction and low cost).

[0105] In some embodiments, the backpropagation and adjustment of the diagnostic confidence parameters of the process diagnostic model and the dynamic response parameters of the digital twin model includes:

[0106] Based on the strategy execution performance index, the overall loss value of model parameter adjustment is calculated. The overall loss value is composed of a weighted average of the quality prediction error term, the energy consumption prediction error term, and the diagnostic accuracy error term.

[0107] The gradient of the dynamic response parameters of each actuator and environmental control device in the digital twin model is calculated based on the overall loss value, and the dynamic response parameters are updated using an adaptive learning rate algorithm according to the gradient direction and magnitude.

[0108] The control strategy scheme is re-simulated based on the updated digital twin model to generate corrected simulation data, and the confidence deviation of each inference node of the process diagnostic model in the semantic diagnostic description is calculated based on the simulation data.

[0109] Based on the confidence bias, the feature weights of the cross-modal attention fusion module in the process diagnostic model and the prior weights of the causal relationships related to the defects or tendencies in the process knowledge graph are adjusted by the backpropagation algorithm.

[0110] Quality prediction error term: Quantifies the error of the digital twin model in predicting the final quality characteristics of the coating film (such as film thickness and gloss), which comes directly from the deviation between actual quality data and simulation prediction data. Energy consumption prediction error term: Quantifies the error of the digital twin model in predicting equipment energy consumption, which comes from the deviation between actual energy consumption and simulation prediction energy consumption. Diagnostic accuracy error term: Quantifies the inference error of the process diagnostic model, which needs to be evaluated indirectly (e.g., if the diagnosis is correct but the strategy formulated based on it does not achieve the expected results, or a large deviation between simulation and reality suggests that the initial diagnosis may be inaccurate). Weighted summation of the overall loss value: Assign different weights to the above three errors according to their importance to the overall effectiveness, and sum them to form the overall loss value (L). L = w1 × L_quality prediction + w2 × L_energy consumption prediction + w3 × L_diagnostic accuracy. The overall loss value is a scalar, and the optimization goal is to minimize this overall loss value. For the digital twin model, the parameters describing the dynamic characteristics of each actuator and environmental control equipment need to be adjusted. For example, the thermal time constant of the heater, the heat transfer coefficient, the opening response time of the spray gun solenoid valve, and the flow gain. Using techniques such as automatic differentiation, the gradient (▽θL) of the overall loss value L with respect to each of the above dynamic response parameters in the twin model is calculated. An adaptive learning rate algorithm, such as Adam, is used to dynamically adjust the step size (learning rate) of each parameter update based on the historical information of the gradient, and then update the parameters in the opposite direction of the gradient, which makes the model parameters adjust in the direction of reducing the prediction errors of quality and energy consumption. The updated digital twin model is used to re-simulate the control strategy scheme that was originally implemented. Since the model parameters have been calibrated, the corrected simulation data obtained from this simulation is considered to be closer to the preset data. The process state reflected by the corrected simulation data is compared with the semantic diagnostic description output by the original process diagnostic model. Focus is placed on each inference node in the diagnostic description (such as the inferred defect type, physical phenomenon, root cause). The new confidence level of the current corrected data supporting these inferences is calculated, and the deviation between it and the original confidence level given by the diagnostic model is calculated. For example, the original diagnostic model inferred that the low temperature was the main cause with an 85% confidence level. However, reviewing the calibrated model revealed that even with the temperature adjusted to the standard value, the simulation quality remained unsatisfactory. This indicates that the confidence bias of the inference regarding the low temperature was negative (the actual support was lower than the original estimate). If the diagnostic inference is biased, it's likely due to an unreasonable allocation of attention weights when fusing multimodal data (spectral, acoustic, environmental fields), focusing too much on irrelevant or noisy features while neglecting crucial ones. Using the confidence bias as a loss signal, the neural network parameters in the cross-modal attention fusion module are adjusted via backpropagation. This allows the module to learn to allocate attention more rationally when encountering similar conditions in the future, extracting feature combinations from multi-source data that are more conducive to making correct inferences.Bias in diagnostic inferences may also stem from inaccurate prior weights (strength of causal relationships) in the knowledge graph. For example, the graph might assign a high weight to the cause of particulate defects due to excessive wind speed, but actual data (backtracking through a calibrated twin model) shows that this causal relationship is weak under the current new coating system. Based on the confidence bias, the prior weights of causal edges in the knowledge graph related to the current defect or tendency are adjusted. If a causal path is repeatedly confirmed, its weight is increased; if a causal path is repeatedly disproven, its weight is decreased.

[0111] The above describes a method for autonomous optimization of a painting process in the embodiments of this application. The computer system in the embodiments of this application will be described in detail below in conjunction with the above-mentioned method for autonomous optimization of a painting process.

[0112] Please see Figure 3 This is a schematic diagram of an exemplary hardware structure of a computer system in an embodiment of this application.

[0113] In some embodiments, the computer system 300 includes a computer device, which may be a terminal device. The computer device includes a processor 301, a memory 302, a sensor module 303, a communication module 304, an input device 305, and an output device 306 connected via a system bus. The processor 301 of the computer device provides computing and control capabilities. The memory 302 of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database is used to store data.

[0114] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0115] In some embodiments of this application, a computer-readable storage medium is provided, including instructions that, when executed on the computer system 300, cause the computer system 300 to perform a method for autonomous optimization of a painting process according to an embodiment of this application.

[0116] In some embodiments of this application, a computer program product is also provided, which, when run on a computer system 300, causes the computer system 300 to execute a method for autonomous optimization of a painting process according to an embodiment of this application.

[0117] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit it. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0118] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state drive), etc.

[0119] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as ROM or random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A method for autonomous optimization of a paint spraying process, characterized in that, The method comprises the following steps: real-time acquisition of multi-modal sensing data in the paint spraying process, input of the multi-modal sensing data into a pre-trained process diagnosis model, output of a semantic diagnosis description of the current paint spraying process state, the multi-modal sensing data comprising spectral image data, acoustic emission data and environmental field data, the semantic diagnosis description comprising identification of paint film defects or tendencies and physical cause inference; generation of a control strategy scheme for eliminating the defects or tendencies according to the semantic diagnosis description, workpiece information of a workpiece to be sprayed and target paint film quality standards, the control strategy scheme comprising a set of coordinated adjustment instructions for a plurality of environmental adjustment devices and spraying execution mechanisms; input of the control strategy scheme into a digital twin model for simulation, prediction of a paint film state evolution path within a preset time after execution of the control strategy scheme, and, based on the simulation prediction result, rolling optimization of the set of coordinated adjustment instructions in the control strategy scheme by a dynamic optimizer, the dynamic optimizer aiming to minimize a weighted sum of quality deviation, energy consumption and material consumption, and taking into account constraints of dynamic response characteristics of each execution mechanism to form a final optimized strategy; execution of the final optimized strategy, and, after the execution is completed, calculation of a strategy execution performance index based on actually acquired paint film final quality data and simulation prediction data, and, according to the strategy execution performance index, reverse propagation and adjustment of diagnosis confidence parameters of the process diagnosis model and dynamic response parameters of the digital twin model, the input of the multi-modal sensing data into the pre-trained process diagnosis model and the output of the semantic diagnosis description of the current paint spraying process state specifically comprise: preprocessing of the spectral image data, the acoustic emission data and the environmental field data, the preprocessing comprising timestamp alignment and spatial position registration; pixel-by-pixel spectral analysis of the preprocessed spectral image data, extraction of absorption and reflectance distribution maps in a preset paint film characteristic waveband to obtain visual spectral features, time-frequency analysis of the preprocessed acoustic emission data, extraction of characteristic frequency band energy ratios representing paint drop particle size distribution and collision energy to obtain acoustic vibration features, and spatial interpolation of the preprocessed environmental field data to generate three-dimensional distribution cloud maps of temperature field, pressure field and airflow velocity vector field in the paint spraying room to obtain environmental physical field features; weighted fusion of the visual spectral features, the acoustic vibration features and the environmental physical field features by a cross-modal attention fusion module to generate a deep working condition representation vector; output of the semantic diagnosis description based on the deep working condition representation vector and a trained process knowledge graph, the output of the semantic diagnosis description based on the deep working condition representation vector and the trained process knowledge graph comprises: calculation of similarity between the deep working condition representation vector and feature vectors of all nodes in the process knowledge graph that are predefined, selection of a plurality of nodes with similarity higher than a first threshold value as a candidate related node set, and pre-labeling of node types on nodes in the process knowledge graph, the node types comprising device state, environmental parameter, physical phenomenon and defect type. a first node of which is of a defect type in the candidate related node set is taken as a reasoning starting point, and a search is performed in the process knowledge graph along a reverse direction of a causal relationship edge with a limited step until a root cause node of which is of a device state or an environmental parameter is searched, all nodes and connection edges searched are combined in sequence into one or more directed paths from the root cause node to the first node as candidate causal reasoning paths; for each of the candidate causal reasoning paths, a defect represented by the first node is determined as a specific defect type, and an occurrence position of the defect is determined according to original space position information corresponding to the generated deep working condition feature vector; all second nodes of which are of a physical phenomenon between the root cause node and the first node on the candidate causal reasoning path are sequentially connected to construct a physical cause chain according to a causal relationship order, a logical confidence of the candidate causal reasoning path is calculated based on an average similarity between the deep working condition feature vector and a feature vector of each node on the candidate causal reasoning path and a prior weight of the candidate causal reasoning path in the knowledge graph, and a severity level is calculated through a preset mapping rule according to the logical confidence and a predefined defect basic severity of the first node; from all the candidate causal reasoning paths, a path with the highest logical confidence is selected, and a generated specific defect type, an occurrence position, a physical cause chain and a severity level are filled into a predefined natural language template to generate the semantic diagnosis description.

2. The method of autonomous optimization of a paint spraying process according to claim 1, characterized in that, generating a control strategy scheme for eliminating the defect or the tendency according to the semantic diagnosis description, workpiece information of a workpiece to be sprayed and a target paint film quality standard includes: analyzing the physical cause chain in the semantic diagnosis description, identifying a starting node of the physical cause chain as a cause node, and identifying a node directly associated with the defect or the tendency as a direct impact factor node; from a preset process countermeasure rule library, a set of corrective actions is matched for the cause node and the direct impact factor node respectively, and the corrective actions correspond to single parameter adjustment of a single environmental regulation device or a spraying execution mechanism; based on geometric features and material properties in the workpiece information and index weights in the target paint film quality standard, all the matched corrective actions are prioritized and conflict detected; according to the prioritization and conflict detection results, a strategy skeleton including multiple candidate atomic action sequences is generated, and the control strategy scheme is generated based on the strategy skeleton.

3. The method of autonomous optimization of a paint spraying process according to claim 2, characterized in that, the control strategy scheme is generated based on the strategy skeleton, which includes: a corresponding device response model is called to predict a dynamic change curve of a target process parameter after each candidate atomic action sequence in the strategy skeleton is executed; the dynamic change curve is compared with an ideal process parameter trajectory corresponding to the target paint film quality standard to calculate an expected performance score of each candidate atomic action sequence; selecting a target candidate atomic action sequence with the highest expected performance score as a basis, and converting each atomic action in the target candidate atomic action sequence into an execution instruction with an execution time point, a duration, a target value, and a change rate according to a rolling optimization requirement of the dynamic optimizer; sequencing all the execution instructions in chronological order, embedding inter-device synchronization and interlocking logic, and encapsulating to form the control strategy scheme.

4. The method of autonomous optimization of a paint spraying process of claim 1, wherein, The strategy execution performance index is calculated based on the actually obtained paint film final quality data and the simulation prediction data, and specifically includes: extracting the same set of quality characteristics from the paint film final quality data and the simulation prediction data, the quality characteristics including average film thickness, film thickness uniformity, gloss, color difference, and area proportion of defect area; for each quality characteristic, calculating the deviation between the actual value and the simulation prediction value, and normalizing the deviation to the corresponding tolerance range to obtain the single-item deviation degree of the quality characteristic; according to the weight coefficients of each quality characteristic pre-set in the target paint film quality standard, performing weighted summation on the single-item deviation degrees to obtain the overall quality coincidence degree index; combining the energy and material cost data consumed in executing the final optimization strategy, comprehensively operating the overall quality coincidence degree index and the unit cost to generate the strategy execution performance index.

5. The method of autonomous optimization of a paint spraying process of claim 1, wherein, The reverse propagation and adjustment of the diagnosis confidence parameters of the process diagnosis model and the dynamic response parameters of the digital twin model include: based on the strategy execution performance index, calculating the overall loss value of model parameter adjustment, which is composed of quality prediction error, energy consumption prediction error, and diagnosis accuracy error; based on the overall loss value, calculating the gradient of the dynamic response parameters of each execution mechanism and environmental regulation device in the digital twin model, and updating the dynamic response parameters using an adaptive learning rate algorithm according to the gradient direction and amplitude; based on the updated digital twin model, re-simulating the control strategy scheme to generate corrected simulation data, and calculating the confidence deviation of each inference node in the semantic diagnosis description based on the simulation data; according to the confidence deviation, adjusting the feature weight of the cross-modal attention fusion module in the process diagnosis model and the prior weight of the causal relationship related to the defect or tendency in the process knowledge graph through a back propagation algorithm.

6. A computer system comprising a memory, a processor, and a computer program stored on the memory, wherein the computer program comprises instructions that, when executed by the processor, cause the processor to perform the method of any one of claims 1 to 5. The processor executes the computer program to implement the steps of the method of any one of claims 1-5.

7. A computer readable storage medium having stored thereon computer programs / instructions, characterized in that, The computer program / instructions are executed by the processor to implement the steps of the method of any one of claims 1-5.

8. A computer program product comprising computer programs / instructions, characterized in that, The computer program / instructions are executed by the processor to implement the steps of the method of any one of claims 1-5.

Images