Spraying quality regulation and control method and system based on fuzzy comprehensive evaluation

Abstract

The invention provides a spraying quality regulation and control method and system based on fuzzy comprehensive evaluation, and relates to the technical field of artificial intelligence and industrial detection.The method comprises the steps that a spraying surface image collected in real time is processed through an improved lightweight target detection model, the type of a spraying area is recognized, and visual feature indexes are extracted; inputting the indexes into a pre-constructed fuzzy comprehensive evaluation model, and outputting a'good ', 'general' or'bad 'comprehensive quality grade in combination with an evaluation standard of dynamic calibration and a weight determined by principal component analysis; and according to the evaluation result and the number of times of spraying of the current area, a control instruction of finishing spraying, marking supplementary spraying or re-spraying is generated and fed back to the robot to be executed, and closed-loop regulation and control are achieved. According to the method, deep learning and multi-criterion decision are fused, the accuracy and real-time performance of spraying quality evaluation are improved, the problems that a traditional method is low in efficiency and poor in adaptability are solved, and the spraying quality is effectively improved.

Description

Technical Field

[0001] This invention relates to the technical field of artificial intelligence and industrial inspection, and more specifically, to a method and system for controlling the quality of spraying based on fuzzy comprehensive evaluation. Background Technology

[0002] As a key supporting structure of wind turbine generators, wind turbine towers are exposed to harsh natural environments for extended periods. The quality of their surface anti-corrosion coating directly affects the tower's service life and overall structural safety. In recent years, automated equipment such as wall-climbing robots have been introduced into the application of anti-corrosion coatings for these structures to replace traditional manual high-altitude work and improve the level of construction automation.

[0003] To achieve autonomous closed-loop control of the spraying operation, real-time online evaluation of the spraying effect is required, and the robot's behavior needs to be adjusted immediately based on the evaluation results. However, this goal faces significant challenges in practice. On the one hand, evaluating the spraying effect requires a comprehensive assessment of multiple indicators, which are often difficult to quantify precisely using traditional threshold judgment methods, leading to discrepancies between the evaluation results and the actual complex quality state. On the other hand, the mobile computing platform carried by the wall-climbing robot is severely limited in terms of computing power, storage space, and power consumption. This makes it difficult to directly deploy and apply many computationally complex advanced evaluation algorithms, failing to meet the real-time requirements of the spraying operation process for detection and decision-making speed. Existing technical solutions typically separate quality inspection and operation control into two relatively independent stages, or use relatively simple and fixed rules for decision-making, making it difficult to achieve accurate and adaptive real-time evaluation and control under resource-constrained conditions.

[0004] Existing technologies for real-time evaluation of spraying effects on mobile robot platforms suffer from insufficient evaluation efficiency, accuracy, and adaptability, making it difficult to effectively support intelligent operation decision-making. Summary of the Invention

[0005] The purpose of this invention is to provide a spraying quality control method and system based on fuzzy comprehensive evaluation, so as to alleviate the technical problems of low efficiency and poor adaptability in the prior art.

[0006] In a first aspect, embodiments of the present invention provide a method for controlling the spraying quality based on fuzzy comprehensive evaluation, comprising: processing real-time acquired images of the sprayed surface using an improved lightweight target detection model to identify the category of the sprayed area; the improved lightweight target detection model is built based on YOLOv5s and includes a bidirectional feature fusion path; generating visual feature indicators of the spraying quality based on the category identification results of the sprayed area; inputting the visual feature indicators into a pre-constructed fuzzy comprehensive evaluation model to generate a comprehensive quality evaluation level corresponding to the current sprayed area; the fuzzy comprehensive evaluation model is constructed based on dynamically calibrated evaluation criteria and indicator weights determined through principal component analysis; generating corresponding control instructions based on the comprehensive quality evaluation level and the number of times the current area has been sprayed and sending them to the robot for execution; the control instructions include ending the spraying of the current area, marking the position of the current area for subsequent re-spraying, or re-spraying the current area.

[0007] In some optional implementations, the improved lightweight object detection model also includes a feature enhancement module that integrates multi-scale convolution and residual connections. The feature enhancement module extracts multi-scale features by setting convolution kernels of different sizes in parallel, and then concatenates the extracted multi-scale features with the original input of the feature enhancement module before outputting the result.

[0008] In some optional implementations, the aforementioned visual characteristics include coating uniformity, integrity, and thickness.

[0009] In some optional implementations, the method for establishing the evaluation criteria for dynamic calibration includes: constructing an initial standard evaluation matrix of the standard values ​​of each visual feature index under each of the above comprehensive quality evaluation levels; and dynamically calibrating the standard values ​​in the initial standard evaluation matrix based on the statistical distribution of each visual feature index in the training sample data to obtain the corrected evaluation matrix.

[0010] In some optional implementations, the weights of the indicators are determined through principal component analysis, including: constructing a correlation coefficient matrix for the standardized visual feature indicator sample data; solving for the eigenvalues ​​and corresponding eigenvectors of the correlation coefficient matrix; selecting the top k principal components whose cumulative variance contribution rate reaches a preset threshold, and using the corresponding eigenvectors as new comprehensive evaluation factors; and determining the normalized variance contribution rate of each principal component as the weight of the corresponding common factor in the comprehensive evaluation.

[0011] In some optional implementations, control instructions are generated based on the overall quality evaluation level and the number of times the current area has been painted, including: if the overall quality evaluation level is "very good", then an instruction to end the painting of the current area is generated; if the overall quality evaluation level is "poor", then an instruction to paint the current area again is generated; if the overall quality evaluation level is "average", then the number of times the current area has been painted is determined: when the number of times the painting is less than or equal to two, an instruction to mark the position for subsequent repainting is generated; when the number of times the painting is three or more, an instruction to paint the current area again is generated.

[0012] In some optional implementations, the improved lightweight object detection model described above uses the EIOU loss function as the loss function for the bounding box regression task during the training phase. The improved lightweight object detection model also includes: a Focus module, which is used to slice and reconstruct the input image to reconstruct the spatial dimension information to the channel dimension; and a spatial pyramid pooling module, which is used to extract multi-scale features and perform splicing and fusion by performing max pooling operations of different scales in parallel.

[0013] Secondly, embodiments of the present invention provide a spraying quality control system based on fuzzy comprehensive evaluation, comprising: a spraying identification and classification module, used to process real-time acquired sprayed surface images using an improved lightweight target detection model to identify the category of the sprayed area; the improved lightweight target detection model is built based on YOLOv5s and includes a bidirectional feature fusion path; a feature index determination module, used to generate visual feature indices of spraying quality based on the category identification results of the sprayed area; and a fuzzy comprehensive evaluation module, used to input the visual feature indices into a pre-constructed fuzzy comprehensive evaluation model to generate a comprehensive quality evaluation level corresponding to the current sprayed area; the fuzzy comprehensive evaluation model is constructed based on dynamically calibrated evaluation criteria and index weights determined through principal component analysis. The decision control module is used to generate corresponding control instructions based on the above comprehensive quality evaluation level and the number of times the current area has been sprayed, and send them to the robot for execution. The control instructions include ending the spraying of the current area, marking the position of the current area for subsequent re-spraying, or spraying the current area again.

[0014] Thirdly, embodiments of the present invention provide an electronic device, including a memory and a processor, wherein the memory stores a computer program that can run on the processor, and the processor executes the computer program to implement the steps of the method described in any of the first aspects above.

[0015] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing computer-executable instructions, which, when invoked and executed by a processor, cause the processor to perform the method described in any of the first aspects above.

[0016] This invention provides a method and system for controlling spray coating quality based on fuzzy comprehensive evaluation. The method processes real-time acquired images of the sprayed surface using an improved lightweight target detection model, identifies the sprayed area category, and extracts visual feature indicators. These indicators are input into a pre-constructed fuzzy comprehensive evaluation model, which, combined with dynamically calibrated evaluation criteria and weights determined by principal component analysis, outputs a comprehensive quality level of "very good," "average," or "poor." Based on the evaluation results and the number of times the current area has been sprayed, control commands are generated to end spraying, mark for respraying, or re-spray, and these commands are fed back to the robot for execution, achieving closed-loop control. This invention integrates deep learning and multi-criteria decision-making, improving the accuracy and real-time performance of spray coating quality assessment, solving the problems of low efficiency and poor adaptability of traditional methods, and effectively improving spray coating quality. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments of the present invention will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating a spraying quality control method based on fuzzy comprehensive evaluation provided in an embodiment of the present invention; Figure 2 A schematic diagram of a spraying quality control system based on fuzzy comprehensive evaluation provided in an embodiment of the present invention; Figure 3 A flowchart illustrating another spraying quality control method based on fuzzy comprehensive evaluation provided in this embodiment of the invention; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] To achieve autonomous closed-loop control of the spraying operation, real-time online evaluation of the spraying effect is required, and the robot's operational behavior needs to be adjusted immediately based on the evaluation results. Currently, the operational decisions of wall-climbing robots still mainly rely on human intervention, and a comprehensive performance evaluation system has not yet been established. Furthermore, since the spraying quality evaluation needs to be performed synchronously with the robot's operation, more stringent requirements are placed on the real-time performance of the detection system.

[0021] Based on this, the present invention provides a spraying quality control method based on fuzzy comprehensive evaluation, which can combine deep learning detection with multi-criteria decision model, improve the scientific nature of evaluation while ensuring high detection speed, and is suitable for wind power tower anti-corrosion operation scenarios with high real-time requirements. It improves the accuracy and real-time performance of spraying quality assessment, solves the problems of low efficiency and poor adaptability of traditional methods, and effectively improves spraying quality.

[0022] To facilitate understanding of this embodiment, a detailed description of a spraying quality control method based on fuzzy comprehensive evaluation disclosed in this invention will be provided first. (See [link to relevant documentation]). Figure 1 The diagram shows a flow chart of a spraying quality control method based on fuzzy comprehensive evaluation. This method can be executed by electronic equipment and mainly includes the following steps S102 to S106: Step S102: The real-time acquired sprayed surface image is processed using an improved lightweight target detection model to identify the category of the sprayed area; Before step S102 above, images of the surface to be sprayed can be acquired in real time during the wall-climbing robot's spraying operation. One application scenario of this invention is that the wall-climbing robot performs anti-corrosion coating spraying operations on the surface of a wind turbine tower. To ensure real-time performance, images can be acquired in real time to achieve real-time monitoring of the spraying effect. Specifically, an improved lightweight target detection model can be used to process the sprayed surface images to classify the spraying conditions.

[0023] In this embodiment, the improved lightweight target detection model is built on YOLOv5s and includes a bidirectional feature fusion path.

[0024] Preferably, the improved lightweight target detection model includes a bidirectional feature fusion path, which can be composed of a combination of a top-down FPN path and a bottom-up PAN path; wherein, the FPN path is used to pass high-level semantic information to low-level feature maps to enhance the semantic expressive power of shallow features, and the PAN path is used to pass low-level detailed information back to high-level feature maps to enhance localization accuracy.

[0025] In another embodiment, the improved lightweight object detection model further includes a feature enhancement module that integrates multi-scale convolution and residual connections; the feature enhancement module extracts multi-scale features by setting convolution kernels of different sizes in parallel, and outputs the extracted multi-scale features after concatenating them with the original input of the feature enhancement module.

[0026] As a concrete example, the improved lightweight object detection model is based on the YOLOv5s architecture, and its backbone network embeds a feature enhancement module that integrates the Inception multi-scale convolutional structure and the ResNet residual connection mechanism.

[0027] Preferably, the feature enhancement module is configured as follows: it extracts multi-scale information of the input feature map in parallel using 1×1, 3×3, and 5×5 convolutional kernels, and concatenates the outputs of each branch and adds them to the identity mapping to form an enhanced output feature map, which serves as the input to the subsequent feature pyramid network.

[0028] In another embodiment, the improved lightweight object detection model uses the EIOU loss function as the loss function for the bounding box regression task during the training phase. The EIOU loss function simultaneously considers the overlap area between the predicted and ground truth boxes, the center point distance deviation, and the aspect ratio difference, and sets independent penalty terms for each for joint optimization. The improved lightweight object detection model also includes a Focus module used in the input processing stage, which slices and reconstructs the input image to reconstruct spatial dimension information to the channel dimension, so as to achieve spatial compression without information loss.

[0029] The Focus module in the above embodiments is not a traditional attention mechanism, but rather a spatial slicing and reconstruction structure in YOLOv5 used to replace the first convolutional layer. Its core operation is to downsample and slice the input image according to odd and even coordinates (e.g., dividing an RGB image with height H, width W, and 3 channels into four feature maps with height H / 2 and width W / 2), and then concatenating them along the channel dimension to form a high-dimensional feature map with height H / 2, width W / 2, and 4×3 channels. This operation can achieve spatial-to-channel information reconstruction without losing information, reducing the initial computational load.

[0030] The improved lightweight object detection model also includes a Spatial Pyramid Pooling Module (SPP) used in the feature extraction stage, which extracts multi-scale features and performs splicing and fusion by performing max pooling operations of different scales in parallel.

[0031] The SPP module can extract multi-receptive field responses from feature maps by using multiple max pooling operations of different scales (such as M×M, N×N, K×K, and L×L, where M>N>K>L) in parallel, and then concatenate the output with the original feature map. This can effectively enhance the model's ability to perceive targets of different scales. In YOLOv5, it is mainly used to improve the robustness of recognizing multi-scale targets such as spraying defects in complex backgrounds.

[0032] In the above embodiments, the spraying area can be categorized into unsprayed areas, normal sprayed areas, and abnormal sprayed areas, with different categories corresponding to different surface state characteristics and processing strategies. Unsprayed areas refer to the original metal surface that has not yet been covered with paint. Visually, they appear as exposed substrate material with uniform color and no coating texture. In the target detection model, they are identified as the starting area for initiating spraying operations. Normal sprayed areas refer to areas that have undergone qualified spraying. The surface coating is continuous, with consistent color and no obvious defects, visually presenting a uniform coverage effect. These can serve as a benchmark reference area for quality evaluation. Abnormal sprayed areas refer to areas with spraying quality problems, including defects such as missed spraying, runs, orange peel, uneven thickness, or contamination. Their visual characteristics include local color deviation, disordered texture, blurred boundaries, or foreign matter adhesion. Depending on the specific defect type, respraying or reprocessing mechanisms need to be triggered.

[0033] By classifying and identifying these three types of regions, we can accurately locate the regions to be processed and extract the visual feature indicators of the corresponding regions, providing a data foundation for subsequent quality judgment based on fuzzy comprehensive evaluation, and ensuring the pertinence and accuracy of regulatory decisions.

[0034] Step S104: Based on the category recognition results of the sprayed area, generate visual feature indicators of the spraying quality.

[0035] Preferably, visual characteristic indicators include coating uniformity, coating integrity, and coating thickness.

[0036] The uniformity of the coating can be quantified using image grayscale standard deviation, color distribution entropy, or texture consistency algorithms. It reflects the surface smoothness of the coating under light reflection characteristics; a lower value indicates more uniform spraying. The integrity of the coating can be obtained through region connectivity analysis and edge closure calculation in the target detection segmentation results. This measures whether there are missed areas, gaps, or local defects, and is usually expressed as the percentage of effective coverage area or defect pixel density. The coating thickness can be estimated from monocular images using multispectral imaging, laser triangulation, or shadow-based deep learning regression models. The output is a relative or absolute thickness value, mapped to a preset quality level range. These three indicators together constitute the core measurable metrics for spraying quality, supporting subsequent multi-criteria comprehensive evaluation.

[0037] In another embodiment, visual feature indicators may also include: adhesion predictor (based on surface microstructure change analysis), gloss (extracted statistically from the intensity of highlight areas), color difference (CIE-Lab spatial deviation relative to a standard sample), and edge transition smoothness (used to evaluate the coating joint quality in overlapping areas). These indicators can be flexibly introduced according to specific process requirements and sensor configurations to further enhance the evaluation model's discrimination ability and robustness under complex working conditions.

[0038] All of the above indicators can be normalized and then input into the fuzzy comprehensive evaluation model to ensure the rationality and comparability of data with different dimensions during the fusion process.

[0039] Step S106: Input the visual feature indicators into the pre-built fuzzy comprehensive evaluation model to generate the comprehensive quality evaluation level corresponding to the current spraying area; The fuzzy comprehensive evaluation model is constructed based on the evaluation criteria of dynamic calibration and the index weights determined through principal component analysis.

[0040] In one embodiment, the method for establishing the evaluation criteria for dynamic calibration may include: constructing an initial standard evaluation matrix of standard values ​​for each visual feature index under each comprehensive quality evaluation level; and dynamically calibrating the standard values ​​in the initial standard evaluation matrix (e.g., adjusting for normal distribution fitting) based on the statistical distribution of each visual feature index in the training sample data to obtain the corrected evaluation matrix.

[0041] The statistical distribution of each visual feature index in the training sample data may include: obtaining a dataset of qualified area samples from historical spraying operations as training sample data, and calculating the mean and variance of each visual feature index in the sample dataset.

[0042] In one embodiment, determining the index weights through principal component analysis may include: constructing a correlation coefficient matrix from standardized visual feature index sample data; solving for the eigenvalues ​​and corresponding eigenvectors of the correlation coefficient matrix; selecting the top k principal components whose cumulative variance contribution rate reaches a preset threshold, and using the corresponding eigenvectors as new comprehensive evaluation factors; and determining the normalized variance contribution rate of each principal component as the weight of the corresponding common factor in the comprehensive evaluation.

[0043] As a concrete example, the factor set U consists of visual characteristic indicators (coating uniformity, integrity, and thickness). Uniformity, integrity, thickness Then, for each indicator, an evaluation set V containing three levels: "very good", "average", and "poor" is constructed. An initial standard evaluation matrix E is established for each indicator at different levels. This matrix defines the ideal numerical range or threshold that the indicator should achieve at each evaluation level. , E is the evaluation matrix. For evaluation level Lowering the target The standard value.

[0044] To overcome potential discrepancies between standard values ​​and actual sampled data distribution, and to improve the model's applicability, the standard values ​​in the initial standard evaluation matrix E can be dynamically calibrated based on the training sample data. Specifically, the mean of each indicator in the training samples can be calculated. With variance The sample mean reflects the index in the measured data. The average performance level is represented by the standard value, while the variance quantifies its fluctuation. Based on this statistical information, the standard values ​​in the initial standard evaluation matrix E are determined. Dynamic adjustments and calibrations are performed to better reflect the probability distribution of the actual data, resulting in the corrected evaluation matrix E′.

[0045] To eliminate potential multicollinearity among evaluation indicators and extract core information, the sample data was standardized to obtain a standardized matrix X. Subsequently, its correlation coefficient matrix was calculated. By solving for the p eigenvalues ​​of R and their corresponding eigenvectors The purpose of principal component analysis is to transform multiple related indicators into a few unrelated common factors.

[0046] Perform principal component analysis on the sample data of visual feature indicators, extract the top k principal components with a relatively high cumulative contribution rate, and construct a weighted comprehensive evaluation function with their normalized variance contribution rates as weights. According to the magnitudes of the eigenvalues (usually following the principle of selecting a cumulative contribution rate greater than 85%), select the top k principal components as the key common factors. Calculate the variance contribution rates of these common factors , and normalize them to obtain the weight vectors for each factor . Finally, through weighted synthesis, map and reduce the original three-index evaluation set U to a new and comprehensive evaluation grade set , where . After standardizing V, obtain the final standardized evaluation set V' for quality determination, which serves as the direct basis for the spraying decision

[0047] Specifically, the method for generating the comprehensive quality evaluation grades ("very good", "bad", "average") based on the standardized evaluation set V' may include: comparing the standardized comprehensive evaluation value V' with a pre-set double-threshold interval, and determining the final quality grade through the threshold discrimination method. Let the lower limit of the comprehensive evaluation value threshold corresponding to the "very good" grade be Th, and the upper limit corresponding to the "bad" grade be Tl, and satisfy Tl < Th. Then the decision rules are as follows If V' ≥ Th, it is determined that the comprehensive quality grade of the current spraying area is "very good", indicating that all indicators of the coating perform excellently as a whole and meet the acceptance standards; if V' ≤ Tl, it is determined as "bad", indicating that there are obvious quality defects and immediate intervention is required; if Tl < V' < Th, it is determined as "average", indicating that the quality is in a critical state, and there may be local non-compliance but it has not seriously affected the overall performance

[0048] The above double thresholds Th and Tl are not fixed empirical values, but are dynamically calibrated based on the weight vectors obtained from principal component analysis and the statistical distributions (such as mean ± 1 times / 2 times standard deviation) of the measured data of each grade in the training samples, and are calibrated in combination with process requirements and expert scoring results to ensure their adaptability under different working conditions and material systems. In addition, to improve the decision-making stability, the system can introduce a sliding time window mechanism to perform weighted averaging or trend analysis on the V' values collected continuously for multiple times, avoiding misjudgment caused by single measurement noise. This method realizes a clear mapping from the multi-dimensional index fusion result to the clear quality grade, taking into account both scientificity and engineering practicality, and provides a reliable input for the subsequent closed-loop control strategy based on the grade and the number of spraying times

[0049] Step S108, generate corresponding control instructions according to the comprehensive quality evaluation grade and the number of sprayed times in the current area, and send them to the robot for execution The control commands include ending the current area spraying, marking the current area position for subsequent touch-up spraying, or spraying the current area again.

[0050] In one embodiment, the number of times the current area has been sprayed is determined by: establishing a unique identifier for each sprayed area and counting the spraying operations based on a method combining visual recognition results and spatial coordinate mapping. Specifically, this may include the following steps: First, in the initialization phase, based on the 3D model of the structure to be sprayed (such as a wind turbine tower) or the on-site scanned point cloud data, the entire work surface is divided into multiple locatable sub-regions, and a unique geographic label is assigned to each sub-region. This label is mapped to the image pixel coordinates and robot pose information (obtained through IMU, encoder or multi-sensor fusion positioning system). Secondly, during each spraying operation, once the improved lightweight target detection model identifies a certain spraying area category, it combines the real-time pose information of the current wall-climbing robot with the camera's intrinsic and extrinsic parameters, and uses a back-projection algorithm to locate the detection box in the image to the global three-dimensional coordinate system, matching its sub-region label. Then, the system queries the local spraying history database for the number of times the sub-area label has been sprayed. If this is the first time the area has been detected, the initial spraying count is set to 1; otherwise, the count is incremented by 1 and the database is updated. Simultaneously, each quality evaluation result ("Excellent", "Average", "Poor") is also associated with and stored with that spraying record for subsequent analysis and model optimization.

[0051] Furthermore, to improve the accuracy of region consistency matching under cross-view and local occlusion conditions, a region fingerprinting mechanism based on feature descriptors (such as SURF or deep local features) can be introduced: extract the surrounding contextual texture features of each sub-region as a unique "visual fingerprint", and enhance the accuracy of region re-identification by comparing the visual fingerprint in subsequent identification, avoiding false counts caused by repeated observations.

[0052] In addition, the control system can be set with an anti-shake mechanism, which stipulates that multiple detections of the same area within the same time window (such as 30 seconds) will only trigger one count update, preventing repeated accumulation caused by the robot's short pauses or continuous frame detection.

[0053] This method enables precise tracking of the frequency of spraying in each area, ensuring that the decision control module can make reasonable judgments based on the actual cumulative number of sprayings. It is a key supporting link for realizing the differentiated strategy of "marking and respraying" and "re-spraying".

[0054] In one embodiment, a control instruction is generated based on the overall quality evaluation level and the number of times the current area has been painted, including: if the overall quality evaluation level is "very good", then generating an instruction to end the painting of the current area; if the overall quality evaluation level is "poor", then generating an instruction to paint the current area again; if the overall quality evaluation level is "average", then determining the number of times the current area has been painted: if the number of paintings is less than or equal to two, generating an instruction to mark the position for subsequent touch-up painting; if the number of paintings is three or more, generating an instruction to paint the current area again.

[0055] This invention provides a spraying quality control method based on fuzzy comprehensive evaluation. It processes real-time acquired images of the sprayed surface using an improved lightweight target detection model, identifies sprayed area categories, and extracts visual feature indicators. These indicators are input into a pre-constructed fuzzy comprehensive evaluation model, which, combined with dynamically calibrated evaluation criteria and weights determined by principal component analysis, outputs a comprehensive quality level of "very good," "average," or "poor." Based on the evaluation results and the number of times the current area has been sprayed, control commands are generated to end spraying, mark for respraying, or re-spray, and fed back to the robot for execution, achieving closed-loop control. This invention integrates deep learning and multi-criteria decision-making, improving the accuracy and real-time performance of spraying quality assessment, solving the problems of low efficiency and poor adaptability of traditional methods, and effectively improving spraying quality. This invention combines deep learning detection with a multi-criteria decision-making model, ensuring high detection speed while improving the scientific rigor of the evaluation, making it suitable for wind turbine tower corrosion protection operations with high real-time requirements.

[0056] Based on the same inventive concept, this invention also provides a spraying quality control system based on fuzzy comprehensive evaluation, see [link to relevant documentation]. Figure 2 As shown, the system mainly includes the following parts: The spraying identification and classification module 210 is used to process the real-time acquired sprayed surface image using an improved lightweight target detection model to identify the category of the sprayed area; the improved lightweight target detection model is built based on YOLOv5s and includes a bidirectional feature fusion path. The feature index determination module 220 is used to generate visual feature indexes of spraying quality based on the category recognition results of the sprayed area. The fuzzy comprehensive evaluation module 230 is used to input visual feature indicators into a pre-built fuzzy comprehensive evaluation model to generate a comprehensive quality evaluation level corresponding to the current spraying area; the fuzzy comprehensive evaluation model is constructed based on dynamically calibrated evaluation criteria and indicator weights determined through principal component analysis; The decision control module 240 is used to generate corresponding control instructions based on the comprehensive quality evaluation level and the number of times the current area has been sprayed, and send them to the robot for execution. The control instructions include ending the spraying of the current area, marking the position of the current area for subsequent re-spraying, or spraying the current area again.

[0057] The fuzzy comprehensive evaluation-based spraying quality control system provided in this invention embodiment can be specific hardware on the device or software or firmware installed on the device. The device provided in this invention embodiment has the same implementation principle and technical effects as the aforementioned method embodiment. For the sake of brevity, any parts not mentioned in the device embodiment can be referred to the corresponding content in the aforementioned method embodiment. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can all be referred to the corresponding processes in the above method embodiments, and will not be repeated here.

[0058] To facilitate understanding, this invention also provides an application example of a spraying quality control method based on fuzzy comprehensive evaluation, see [link to relevant documentation]. Figure 3 The flowchart of another spraying quality control method based on fuzzy comprehensive evaluation is shown, including: the input image is processed by the improved YOLOv5s model and the fuzzy comprehensive evaluation model respectively, and then enters the decision control module; according to the judgment result, the corresponding control instructions are generated: if the evaluation is "very good", then "end spraying"; if the evaluation is "average" and it is the first or second spraying, then "point marking" is performed; if the evaluation is "poor" or "average" and it is the third spraying, then "re-spraying" is performed.

[0059] Specifically, the key processes of this method mainly include: (1) real-time monitoring of spraying and (2) fuzzy comprehensive evaluation.

[0060] (1) Real-time monitoring of spraying: To ensure real-time judgment, the lightweight YOLOv5s framework is used to classify the spraying situation: <1> Lightweight input processing: The input image is first sliced ​​and reconstructed by the Focus module. The spatial dimension information is reconstructed to the channel dimension by sampling at pixel intervals, forming a feature layer with four channels that are multiplied by several times and halved in size.

[0061] This operation significantly reduces computation and memory usage while avoiding information loss, thus building an efficient data foundation for subsequent feature extraction.

[0062] <2> Multi-scale feature extraction: The SPP module is introduced to extract feature responses with multiple receptive fields by performing four different scale max pooling operations in parallel.

[0063] After multi-scale features are spliced ​​and fused, the model's ability to adapt to changes in the scale of the spraying target is effectively enhanced, and the robustness of identification of defects or regions of different sizes is improved.

[0064] <3> Bidirectional feature fusion mechanism: In the feature pyramid part, three layers of features obtained by downsampling 3, 4 and 5 times in the backbone network are selected to construct a bidirectional fusion path combining FPN and PAN.

[0065] Among them, FPN achieves top-down semantic enhancement, improving the expressive power of shallow features; PAN conveys positional detail information in a bottom-up manner. The two work together to effectively achieve deep fusion of multi-level features, significantly enhancing the model's representation ability in complex spraying backgrounds.

[0066] <4> Detection head optimization and sample balancing strategies: During the anchor box design phase, the fuzzy K-means algorithm is used to pre-cluster the real bounding boxes, and then the traditional K-means algorithm is used for secondary optimization to generate prior box sizes that better fit the actual distribution of the target.

[0067] The regression loss function adopts EIOU loss, which combines constraints on three aspects: overlapping area, center point distance, and width-height difference, effectively improving positioning accuracy.

[0068] Simultaneously, an adaptive confidence balancing mechanism is introduced to alleviate the problem of imbalanced positive and negative samples. In addition, an InRe feature enhancement module is embedded before each detection layer to strengthen the expressive power of key features.

[0069] <5> Loss function improvement: To further improve training stability and convergence efficiency, the original GIOU loss is replaced with EIOU loss as the bounding box regression target.

[0070] EIOU further decouples the width and height differences based on GIOU, forming a more detailed penalty term. This allows the loss function to more comprehensively characterize the geometric differences between the predicted box and the ground truth box, thereby accelerating model convergence and improving localization accuracy.

[0071] <6> Feature enhancement module: This paper integrates the multi-scale convolution concept of Inception with the residual connection mechanism of ResNet. By using convolution kernels of different sizes in parallel to extract multi-scale features, and combining residual connections to promote gradient flow and information preservation, the paper achieves this goal.

[0072] After embedding this module at key layers of the backbone network, the multi-scale convolutional output is concatenated with the original input to form a fusion feature that combines detailed information and contextual semantics. This fusion serves as a high-quality input to the feature pyramid, providing a highly discriminative feature representation for subsequent multi-scale detection. (2) Fuzzy comprehensive evaluation: To achieve scientific evaluation and intelligent decision-making regarding the painting quality of a wall-climbing robot, a quantitative analysis model based on fuzzy comprehensive evaluation was constructed. This model categorizes painting quality into three levels: "Excellent," "Average," and "Poor," and executes corresponding painting control strategies accordingly: if the evaluation is "Poor," the robot is instructed to immediately perform a second painting; if the evaluation is "Average," and it is the first or second painting, the location is marked, and specific repainting is initiated in subsequent cycles; if the evaluation is "Excellent," the area is deemed to be painted satisfactorily, requiring no further action.

[0073] The specific implementation steps of this fuzzy comprehensive evaluation model are as follows: (S301) Construct an evaluation index system and standard evaluation matrix; First, the system selects "uniformity," "integrity," and "thickness" as core evaluation indicators, which together constitute the factor set U. Uniformity, integrity, thickness For each indicator, a standard value is specified in three levels: "very good", "average", and "poor". This allows us to construct an initial standard evaluation matrix E. This matrix defines the ideal numerical range or threshold that the indicators should achieve at each evaluation level. , E is the evaluation matrix. For evaluation level Lowering the target The standard value.

[0074] (S302) Adjust the evaluation criteria based on the distribution of sample data; To overcome potential discrepancies between standard values ​​and actual sampled data distribution, and to improve the model's applicability, the mean values ​​of each indicator in the training samples are calculated. With variance The sample mean reflects the performance of this indicator in the measured data. The average performance level is represented by the standard value, while the variance quantifies its fluctuation. Based on this statistical information, the standard values ​​in the initial standard evaluation matrix E are... Dynamic adjustments and calibrations are performed to better reflect the probability distribution of actual data, resulting in a corrected evaluation matrix E'.

[0075] (S303) Correlation analysis of indicators and factor dimensionality reduction; To eliminate potential multicollinearity among evaluation indicators and extract core information, the sample data was standardized to obtain a standardized matrix X. Subsequently, its correlation coefficient matrix was calculated. By solving for the p eigenvalues ​​of R and their corresponding eigenvectors Principal component analysis was then performed. This aims to transform the original multiple correlated indicators into a few uncorrelated common factors.

[0076] (S304) Construct a weighted comprehensive evaluation model; Based on the magnitude of the eigenvalues ​​(usually selecting those with a cumulative contribution rate greater than 85%), the top k principal components are selected as key common factors. The variance contribution rate of these common factors is then calculated. And normalized them as the weight vector of each factor. Finally, through weighted synthesis, the original three-indicator evaluation set U is mapped and dimensionality reduced into a completely new, comprehensive set of evaluation levels. ,in, After standardizing V, the final standardized evaluation set V' used for quality judgment is obtained, which serves as the direct basis for spraying decisions.

[0077] This invention employs an improved YOLOv5s approach combined with fuzzy comprehensive evaluation to assess the spraying effect and control the quality of anti-corrosion coatings. To meet the engineering requirements for real-time monitoring of spraying conditions, a lightweight YOLOv5s architecture is used as the foundation. Systematic improvements are made to key aspects such as feature extraction, multi-scale fusion, and loss optimization. This constructs an efficient and reliable spraying state classification model, which integrates deep learning and multi-criteria decision-making to improve the accuracy and real-time performance of spraying quality assessment. This solves the problems of low efficiency and poor adaptability of traditional methods, effectively improving spraying quality.

[0078] Based on the same inventive concept, embodiments of the present invention also provide an electronic device, specifically, the electronic device includes a processor and a storage device; the storage device stores a computer program, and the computer program, when run by the processor, executes the method described in any of the above embodiments.

[0079] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. The electronic device 400 includes: a processor 410, a memory 420, a communication interface 430, and a bus 440. The memory 420 stores machine-readable instructions that can be executed by the processor 410. When the electronic device is running, the processor 410 communicates with the memory 420 through the bus 440. The processor 410 executes the machine-readable instructions to perform the steps of the method described above.

[0080] Specifically, the memory 420 and processor 410 can be general-purpose memory and processor, without any specific limitations. When the processor 410 runs the computer program stored in the memory 420, it can execute the above method.

[0081] Processor 410 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of processor 410 or by instructions in software form. The processor 410 may be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it may also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this invention. The general-purpose processor may be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this invention can be directly manifested as execution by a hardware decoding processor, or execution by a combination of hardware and software modules in the decoding processor. The software module can reside in a mature storage medium in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory 420. The processor 410 reads the information from memory 420 and, in conjunction with its hardware, completes the steps of the above method.

[0082] Corresponding to the above method, this embodiment of the invention also provides a computer-readable storage medium storing machine-executable instructions. When the machine-executable instructions are called and run by a processor, the machine-executable instructions cause the processor to perform the steps of the above method.

[0083] In the embodiments provided by this invention, it should be understood that the disclosed apparatus and method can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0084] Furthermore, the units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0085] Furthermore, the functional modules in the various embodiments of the present invention can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0086] It should be noted that if the functionality is implemented as a software module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0087] In this document, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, without necessarily requiring or implying any such actual relationship or order between these entities or operations.

[0088] The above description is merely an embodiment of the present invention and is not intended to limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for controlling spray coating quality based on fuzzy comprehensive evaluation, characterized in that, include: An improved lightweight target detection model is used to process real-time images of sprayed surfaces to identify the categories of sprayed areas. The improved lightweight target detection model is built on YOLOv5s and includes a bidirectional feature fusion path; Based on the category identification results of the sprayed area, a visual feature index of the spraying quality is generated; The visual feature indicators are input into a pre-constructed fuzzy comprehensive evaluation model to generate a comprehensive quality evaluation level corresponding to the current spraying area; The fuzzy comprehensive evaluation model is constructed based on dynamically calibrated evaluation criteria and index weights determined through principal component analysis. Based on the comprehensive quality evaluation level and the number of times the current area has been sprayed, corresponding control instructions are generated and sent to the robot for execution; the control instructions include ending the spraying of the current area, marking the current area position for subsequent re-spraying, or spraying the current area again.

2. The method according to claim 1, characterized in that, The improved lightweight target detection model also includes a feature enhancement module that integrates multi-scale convolution and residual connections; The feature enhancement module extracts multi-scale features by setting convolution kernels of different sizes in parallel, and then concatenates the extracted multi-scale features with the original input of the feature enhancement module before outputting the result.

3. The method according to claim 1, characterized in that, The visual characteristics include the uniformity, integrity, and thickness of the coating.

4. The method according to claim 1, characterized in that, The method for establishing the evaluation criteria for the dynamic calibration includes: Construct an initial standard evaluation matrix for the standard values ​​of each visual feature index under each of the aforementioned comprehensive quality evaluation levels; Based on the statistical distribution of each visual feature index in the training sample data, the standard values ​​in the initial standard evaluation matrix are dynamically calibrated to obtain the corrected evaluation matrix.

5. The method according to claim 1, characterized in that, Principal component analysis is used to determine the weights of the indicators, including: Construct a correlation coefficient matrix from the standardized visual feature index sample data; Solve for the eigenvalues ​​and corresponding eigenvectors of the correlation coefficient matrix; Select the top k principal components whose cumulative variance contribution rate reaches a preset threshold, and use the corresponding eigenvectors as new comprehensive evaluation factors; The normalized variance contribution rate of each principal component is determined as the weight of the corresponding common factor in the comprehensive evaluation.

6. The method according to claim 1, characterized in that, Based on the comprehensive quality evaluation level and the number of times the current area has been coated, control instructions are generated, including: If the overall quality evaluation level is "very good", then an instruction to end the spraying of the current area is generated; If the overall quality evaluation level is "poor", then an instruction is generated to repaint the current area; If the overall quality evaluation level is "average", then determine the number of times the current area has been sprayed: if the number of sprays is less than or equal to two, generate an instruction to mark the position for subsequent re-spraying; if the number of sprays is three or more, generate an instruction to spray the current area again.

7. The method according to claim 1, characterized in that, The improved lightweight object detection model uses the EIOU loss function as the loss function for the bounding box regression task during the training phase; the improved lightweight object detection model also includes: The Focus module is used to slice and reconstruct the input image, reconstructing spatial dimensional information to the channel dimension; The spatial pyramid pooling module is used to extract multi-scale features and perform splicing and fusion by executing max pooling operations of different scales in parallel.

8. A spraying quality control system based on fuzzy comprehensive evaluation, characterized in that, include: The spraying recognition and classification module is used to process real-time acquired sprayed surface images using an improved lightweight target detection model to identify the category of the sprayed area. The improved lightweight target detection model is built on YOLOv5s and includes a bidirectional feature fusion path; The feature index determination module is used to generate visual feature indexes of the spraying quality based on the category identification results of the sprayed area. The fuzzy comprehensive evaluation module is used to input the visual feature indicators into a pre-constructed fuzzy comprehensive evaluation model to generate a comprehensive quality evaluation level corresponding to the current spraying area. The fuzzy comprehensive evaluation model is constructed based on dynamically calibrated evaluation criteria and index weights determined through principal component analysis. The decision control module is used to generate corresponding control instructions based on the comprehensive quality evaluation level and the number of times the current area has been sprayed, and send them to the robot for execution; the control instructions include ending the spraying of the current area, marking the position of the current area for subsequent re-spraying, or spraying the current area again.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions that, when invoked and executed by a processor, cause the processor to perform the method according to any one of claims 1 to 7.

Images