A surface strengthening process identification and coverage detection method and system

Abstract

This application discloses a method and system for surface strengthening process identification and coverage detection, relating to the field of intelligent surface detection technology. The method involves fluorescently treating the workpiece surface to obtain a fluorescence image of the treated surface. A trained deep learning dual-branch network is then used to output the probability of the strengthening process category and the coverage of the strengthened region. The deep learning dual-branch network includes a feature extraction module, a process discrimination branch network, and a coverage segmentation branch network. The feature extraction module generates a fused feature map from the fluorescence image of the workpiece surface. The process discrimination branch network uses a convolutional block attention mechanism to obtain the strengthening process category probability from the fused feature map. The coverage segmentation branch network uses an encoder-decoder semantic segmentation network to determine the coverage of the strengthened region based on the strengthening process category probability and the fused feature map. This application enables accurate discrimination of surfaces with different strengthening processes and precise quantitative detection of coverage.

Description

Technical Field

[0001] This application relates to the field of surface intelligent detection technology, and in particular to a method and system for surface strengthening process identification and coverage detection. Background Technology

[0002] Surface strengthening technology is a key process for improving the reliability and fatigue life of metal components. Shot peening and laser shock peening are widely used in industrial production. Shot peening introduces a residual compressive stress layer by bombarding the workpiece surface with high-speed projectiles, and is characterized by low cost and strong adaptability. Laser shock peening utilizes high-energy laser-induced plasma shock waves to generate deeper residual compressive stress on the workpiece surface, resulting in good surface integrity and significant strengthening effects. Both processes are widely used in the manufacturing of aero-engine blades and key components, and are supported by clear process selection standards and quality requirements.

[0003] Currently, the identification of surface strengthening processes for metal components still relies on manual experience. Shot peening produces surfaces with randomly distributed pits, ranging in diameter from 0.1mm to 1mm, which can be visually identified. However, laser shock peening produces surfaces with numerous micro-pits around 10µm in diameter, making it impossible to distinguish the strengthened areas with the naked eye. A roughness tester is required to measure these localized areas. In summary, manual inspection of strengthened surfaces suffers from drawbacks such as high subjectivity, low efficiency, and susceptibility to misjudgment, failing to meet the demands of high-volume, high-precision inspection.

[0004] Meanwhile, existing machine vision-based inspection systems only identify and detect the coverage of surfaces with specific strengthening processes, resulting in a limited scope of application and failing to establish a unified recognition framework covering multiple strengthening processes. At the algorithm level, existing systems are mainly based on traditional image processing algorithms from the OpenCV open-source vision library, exhibiting poor adaptability to workpieces with different surface conditions and unable to handle complex texture patterns under different strengthening processes.

[0005] Therefore, there is an urgent need for a visual inspection system that can automatically identify surface strengthening processes and accurately calculate the coverage of the strengthened area in order to cope with mass production in the context of smart manufacturing. Summary of the Invention

[0006] The purpose of this application is to provide a method and system for identifying surface strengthening processes and detecting coverage, which can accurately identify surfaces with different strengthening processes and accurately quantify the coverage.

[0007] To achieve the above objectives, this application provides the following solution.

[0008] In a first aspect, this application provides a method for identifying surface strengthening processes and detecting coverage, including: Fluorescent treatment is applied to the surface of the workpiece; Acquire fluorescence images of the workpiece surface after fluorescence treatment; Based on the fluorescence image of the workpiece surface, a trained deep learning dual-branch network is used to output the enhancement process category probability and enhancement region coverage of the workpiece surface. The deep learning dual-branch network includes a feature extraction module, a process discrimination branch network, and a coverage segmentation branch network. The feature extraction module is used to input the fluorescence image of the workpiece surface and generate a fused feature map through the EfficientNet backbone network and the feature pyramid network. The process discrimination branch network adopts a convolutional block attention mechanism to obtain the enhancement process category probability from the fused feature map. The coverage segmentation branch network adopts a semantic segmentation network with an encoder-decoder architecture, determines the enhancement region segmentation result based on the enhancement process category probability and the fused feature map, and then determines the enhancement region coverage based on the enhancement region segmentation result.

[0009] Optionally, the workpiece surface is subjected to fluorescent treatment, including: The workpiece surface is cleaned using organic solvents; Apply the fluorescent penetrant to the cleaned workpiece surface and allow it to penetrate into the pits on the workpiece surface. The residual fluorescent penetrant on the surface of the workpiece is removed by water washing to obtain a cleaned workpiece; Dry the cleaned workpieces. A white developer is applied to the surface of the dried workpiece using a spraying method, which adsorbs the fluorescent penetrant onto the surface layer of the workpiece.

[0010] Optionally, acquiring a fluorescence image of the workpiece surface after fluorescence treatment includes: Real-time reception of raw fluorescence images of the workpiece surface after fluorescence treatment; The original fluorescence image is preprocessed to obtain a fluorescence image of the workpiece surface after fluorescence processing; the preprocessing includes: nonlocal mean filtering for noise reduction, adaptive histogram equalization for contrast enhancement, geometric correction for distortion elimination, and image normalization.

[0011] Optionally, the feature extraction module includes: a cascaded EfficientNet backbone network and a feature pyramid network; The EfficientNet backbone network is used to extract multi-scale features from the fluorescence image of the workpiece surface; Feature pyramid networks are used to fuse multi-scale features to generate fused feature maps.

[0012] Optionally, the process discrimination branch network includes: a convolutional block attention module, a dual pooling layer, a Dropout layer, and a fully connected layer connected in sequence; The convolutional block attention module is used to perform dual attention enhancement on the fused feature map to obtain the enhanced feature map; The dual pooling layer is used to perform global average pooling and max pooling in parallel on the enhanced feature map and then concatenate them into a global feature vector. After overfitting is suppressed by the Dropout layer, the global feature vector is then output through a fully connected layer to enhance the probability of the process category.

[0013] Optionally, the coverage segmentation branch network includes: an encoder and a decoder; The encoder takes the process-sensitive features formed by channel splicing of enhanced process category probabilities and fused feature maps as input, and uses the void space pyramid pooling module to generate high-level semantic features. The decoder is used to input the high-level semantic features, output the enhanced region segmentation result through transposed convolution upsampling, and then output the enhanced region coverage rate based on the enhanced region segmentation result; the enhanced region segmentation result includes pixel-level background, enhanced region and unenhanced region; the enhanced region coverage rate is equal to the ratio of the number of pixels in the enhanced region to the total number of pixels in the effective region; wherein, the total effective region refers to the region other than the background.

[0014] Optionally, the training process of the deep learning dual-branch network includes: Fluorescence images of the surfaces of multiple samples were acquired; the sample surfaces were covered by three different strengthening processes and unstrengthened surfaces. The acquired fluorescence image is subjected to data augmentation to obtain a data-enhanced fluorescence image; the data augmentation includes elastic transformation, color dithering, and random noise injection; Each data-enhanced fluorescence image is labeled with scene annotation and segmentation region mask marking to obtain a training dataset; the scene annotation includes enhancement process category annotation and no enhancement annotation; the segmentation region includes background, enhanced region and no enhancement region; The loss function of the process discrimination branch network is set to the label smoothing cross-entropy loss function, and the loss function of the coverage segmentation branch network is a weighted loss function composed of cross-entropy loss, Dice loss and boundary Focal loss. Based on the training dataset, the deep learning dual-branch network is trained using the label smooth cross-entropy loss function and the weighted loss function to obtain the trained deep learning dual-branch network.

[0015] Optionally, the process discrimination branch network adopts a transfer learning strategy: the process discrimination branch network is obtained through pre-training before training the deep learning dual-branch network.

[0016] Secondly, this application provides a surface strengthening process identification and coverage detection system, including: a sensing device and a computer; The sensing device is used to acquire fluorescence images of the workpiece surface after fluorescence treatment; The computer uses a trained deep learning dual-branch network to output the enhancement process category probability and enhancement region coverage of the workpiece surface based on the fluorescence image of the workpiece surface. The deep learning dual-branch network includes a feature extraction module, a process discrimination branch network, and a coverage segmentation branch network. The feature extraction module takes the fluorescence image of the workpiece surface as input and generates a fused feature map through an EfficientNet backbone network and a feature pyramid network. The process discrimination branch network uses a convolutional block attention mechanism to obtain the enhancement process category probability from the fused feature map. The coverage segmentation branch network uses a semantic segmentation network with an encoder-decoder architecture to determine the enhancement region segmentation result based on the enhancement process category probability and the fused feature map, and then determines the enhancement region coverage based on the enhancement region segmentation result.

[0017] Optionally, the sensing device includes: an optical bracket, a light source bracket, a mobile platform, an industrial camera, an imaging lens, and an ultraviolet excitation light source; The optical support is a cantilever structure, with an industrial camera mounted at the end of the cantilever; the imaging lens is integrated below the industrial camera. The light source support is mounted on the optical support and located below the imaging lens; an ultraviolet excitation light source is fixed on the light source support, which is used to excite the fluorescence on the surface of the workpiece after fluorescence treatment; The moving platform is located below the light source bracket. The moving platform is used to place the workpiece and move the workpiece. The signal output terminal of the industrial camera is connected to the signal input terminal of the computer; the industrial camera is used to acquire fluorescence images of the workpiece surface after fluorescence excitation through the imaging lens and transmit them to the computer.

[0018] According to the specific embodiments provided in this application, this application has the following technical effects.

[0019] This application provides a method and system for identifying and detecting surface strengthening processes. The method involves fluorescently treating the workpiece surface and introducing a fluorescent penetration process to convert random pits on shot-peened surfaces and micro-pits on laser-shock-strengthened surfaces into high-contrast images. This improves the visual discrimination of different strengthening processes by orders of magnitude and breaks through the technical bottleneck of traditional visible light imaging, which is limited by feature size, reflectivity, and material differences. The method is applicable to the identification and detection of surfaces with different strengthening processes.

[0020] In terms of strengthening process identification, deep learning algorithms are used for the intelligent identification and differentiation of surface features: based on the EfficientNet-B3 backbone network and combined with the convolutional block attention mechanism, the grid automatically focuses on the distribution pattern differences of fluorescent points, accurately capturing the array pattern features of two different strengthening processes, shot peening and laser shock strengthening, on the workpiece surface, and achieving high-precision discrimination of surfaces with different strengthening processes.

[0021] In terms of coverage detection, a semantic segmentation network with an encoder-decoder architecture is adopted to capture multi-scale contextual information in parallel, which significantly improves the pixel-level extraction accuracy of reinforced region boundaries, thereby achieving accurate quantitative detection of coverage. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart illustrating a surface strengthening process identification and coverage detection method provided in an embodiment of this application.

[0024] Figure 2 This is a schematic diagram of the fluorescence imaging process provided in the embodiments of this application.

[0025] Figure 3 This is a schematic diagram of the overall process of a surface strengthening process identification and coverage detection method provided in an embodiment of this application.

[0026] Figure 4 This is a schematic diagram of the hardware structure of a surface strengthening process identification and coverage detection system provided in an embodiment of this application.

[0027] Reference numerals: Optical bracket-1, Industrial camera-2, Imaging lens-3, Camera clamp-4, Focusing slider-5, Ultraviolet excitation source-6, Light source bracket-7, Moving platform-8, Cantilever-9, Stage-10, Computer-11. Detailed Implementation

[0028] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0029] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0030] In one exemplary embodiment, such as Figure 1 As shown, a method for identifying surface strengthening processes and detecting coverage is provided. In this embodiment, the method includes the following steps 101 to 103.

[0031] Step 101: Perform fluorescence treatment on the surface of the workpiece.

[0032] Step 102: Obtain fluorescence images of the workpiece surface after fluorescence treatment.

[0033] Step 103: Based on the fluorescence image of the workpiece surface, the trained deep learning dual-branch network is used to output the enhancement process category probability and enhancement region coverage of the workpiece surface. The deep learning dual-branch network includes a feature extraction module, a process discrimination branch network, and a coverage segmentation branch network. The feature extraction module is used to input the fluorescence image of the workpiece surface and generate a fused feature map through the EfficientNet backbone network and the feature pyramid network. The process discrimination branch network adopts a convolutional block attention mechanism to obtain the enhancement process category probability from the fused feature map. The coverage segmentation branch network adopts a semantic segmentation network with an encoder-decoder architecture, determines the enhancement region segmentation result based on the enhancement process category probability and the fused feature map, and then determines the enhancement region coverage based on the enhancement region segmentation result.

[0034] For example, the workpiece can be a spiral tooth or a bevel gear.

[0035] In another exemplary embodiment of this application, step 101 above performs fluorescence treatment on the workpiece surface, including pre-cleaning, penetration treatment, cleaning removal, drying treatment, and development treatment, to enhance the fluorescence imaging effect of surface pits. Figure 2The fluorescent imaging process flow shown is as follows: Pre-cleaning: Using a high-pressure water gun and an organic solvent (ethanol), oil, metal shavings, and other contaminants are removed from the workpiece surface. After cleaning, the workpiece is dried to remove any remaining moisture. Penetration treatment: A fluorescent penetrant is applied to the workpiece surface and left to stand for 20 to 30 minutes to allow it to fully penetrate into the surface pits. Cleaning removal: Residual fluorescent developer is removed from the surface using a water washing method, taking care to retain the penetrant that has already penetrated. Drying treatment: The cleaned workpiece is placed in a hot air circulating drying room to accelerate the removal of surface moisture. The temperature of the drying room is maintained at 50℃-70℃. The drying time is approximately 20 minutes. Imaging treatment: A thin layer of white developer is applied using a spray method and left to stand for 10 minutes to allow the fluorescent penetrant to be fully absorbed to the surface, forming an enlarged fluorescent indication effect.

[0036] The fluorescence treatment process in step 101 can be summarized as follows: cleaning the workpiece surface with an organic solvent; applying a fluorescent penetrant to the cleaned workpiece surface and allowing the fluorescent penetrant to penetrate into the pits on the workpiece surface; removing the residual fluorescent penetrant from the workpiece surface by water washing to obtain a cleaned workpiece; drying the cleaned workpiece; and applying a white developer to the dried workpiece surface using a spraying method to adsorb the fluorescent penetrant onto the surface layer of the workpiece.

[0037] In another exemplary embodiment of this application, step 102 involves image acquisition and image preprocessing, including denoising, enhancement, and normalization. Step 102 can be implemented as follows: receiving the original fluorescence image of the workpiece surface after fluorescence treatment in real time; preprocessing the original fluorescence image to obtain a fluorescence image of the workpiece surface after fluorescence treatment; the preprocessing includes: non-local mean filtering for denoising, adaptive histogram equalization for contrast enhancement, geometric correction for distortion elimination, and image normalization. Non-local mean filtering suppresses image noise, adaptive histogram equalization enhances the contrast of fluorescence spots, geometric correction eliminates lens distortion, and image normalization inputs the image to a fixed size deep learning network.

[0038] For example, firstly, nonlocal mean filtering is used for noise reduction, with the filtering intensity adaptively adjusted according to the local noise level of the image; then, contrast is enhanced by limiting contrast-limited adaptive histogram equalization (CLAHE), dividing the image into 8×8 sub-regions, with the cropping constraint parameter set to 3.0; based on the radial and tangential distortion coefficients pre-calibrated by Zhang's calibration method, pixel remapping is achieved through bilinear interpolation to eliminate the influence of lens distortion; finally, the image size is uniformly adjusted to 1024×1024 pixels, and after the pixel values ​​are linearly mapped to the [0,1] interval, the global mean and standard deviation are calculated, Z-score standardization is performed, and the standardized image tensor is output.

[0039] In another exemplary embodiment of this application, step 103 above performs deep learning analysis on the fluorescence image of the workpiece surface to achieve process category determination and coverage calculation. The deep learning dual-branch network includes a feature extraction module and two cascaded sub-networks, which include a process discrimination branch network and a coverage segmentation branch network.

[0040] In one example, the feature extraction module includes: a cascaded EfficientNet backbone network and a Feature Pyramid Network; the EfficientNet backbone network is used to extract multi-scale features from the fluorescence image of the workpiece surface; the Feature Pyramid Network (FPN) is used to fuse the multi-scale features to generate a fused feature map.

[0041] In another example, the process discrimination branch network employs a convolutional neural network architecture, embedding a spatial attention mechanism module before the fully connected layer to focus on the distribution pattern features of fluorescent spots. The network input is a fused feature map, and the output is a probability distribution of the process type, including three categories: shot peening enhancement, laser enhancement, and no enhancement. Specifically, the process discrimination branch network includes: a convolutional block attention module, a dual pooling layer, a Dropout layer, and a fully connected layer connected in sequence. The convolutional block attention module performs dual attention enhancement on the fused feature map to obtain an enhanced feature map. The dual pooling layer performs parallel global average pooling and max pooling operations on the enhanced feature map and concatenates them into a global feature vector. After overfitting is suppressed by the Dropout layer, the global feature vector is output as an enhanced process category probability through the fully connected layer.

[0042] As can be seen, the process discrimination branch network adopts an "attention enhancement + dual pooling fusion" classification architecture, embedding a Convolutional Block Attention Module (CBAM). The channel attention submodule performs parallel global average pooling and max pooling operations, generates a weight vector through a shared multilayer perceptron, and then weights it channel by channel. The spatial attention submodule pools the channel attention output along the channel dimension, generates a single-channel weight map through a convolutional layer, and then multiplies it pixel by pixel. After dual attention enhancement, the feature maps are concatenated into a global feature vector through parallel global average and max pooling operations. Dropout suppresses overfitting, and the vector is then input into a two-layer fully connected network, outputting the log-probability of three processes, which is then processed by Softmax to generate a probability distribution. The process discrimination branch network achieves... Figure 3 The function of the process discrimination branch is achieved through global average pooling, global max pooling, and a fully connected network.

[0043] In another example, the coverage segmentation branch network employs a process-aware architecture, concatenating the process discrimination output (a category probability vector composed of the probabilities of all enhanced process categories) with the fused feature map channels to form process-sensitive features. The specific process is as follows: First, the process discrimination branch network outputs the category probability vector of the enhanced process category. and class probability vector Broadcast replication is performed in the spatial dimension, expanding its dimension to... The expanded probabilistic feature map is obtained, where each pixel is associated with the same global process category probability. Subsequently, the expanded probabilistic feature map and the fused feature map are concatenated along the channel dimension to form the process-sensitive feature. This process physically injects abstract process category information into spatial features, enabling the network to adaptively adjust the receptive field and weights for subsequent segmentation based on the identified process type. express A 3D real vector space, express A 3D real vector space, express A 3D real vector space, The number of channels represents the category probability vector. Indicates altitude, Indicates width, This indicates the number of channels in the fused feature map. The encoder uses the Atrous Spatial Pyramid Pooling (ASPP) module to capture multi-scale context with multiple dilation rates. After dimensionality reduction and concatenation, the features are fused via convolution to generate high-level semantic features. Specifically, the encoder receives the aforementioned process-sensitive features as input and captures multi-scale contextual information through the ASPP module. The ASPP module deploys five branches in parallel: one 1×1 convolutional branch for extracting local detail features; three 3×3 atrous convolutional branches with dilation rates of 6, 12, and 18 respectively, to capture pit morphology features at different scales with exponentially growing receptive fields; and one global average pooling branch to extract global contextual information and prevent feature loss. The outputs of these five branches are concatenated along the channel dimension to form a feature map containing rich multi-scale information. Subsequently, dimensionality reduction and channel fusion are performed through a 1×1 convolutional layer to eliminate redundant information and compress the feature dimension, ultimately generating high-level semantic features guided by process discrimination. This advanced semantic feature preserves both the process category attributes and integrates multi-scale spatial context, providing a foundation for the decoder to perform high-precision pixel-level boundary restoration. The decoder upsamples through transposed convolutions, concatenates them with low-level features from the backbone network after dimensionality reduction, refines the boundaries through convolution, upsamples to the input resolution, and outputs the segmentation log odds. The coverage segmentation branch network implements pixel-level segmentation using an encoder-decoder structure and calculates the enhanced coverage. The coverage segmentation branch network adopts a semantic segmentation network with an encoder-decoder architecture, preferably using the DeepLabV3+ structure. The coverage segmentation branch network uses fused feature maps as auxiliary input to achieve fine segmentation guided by classification. It outputs a pixel-level three-class mask: background, enhanced region, and unenhanced region. Based on the segmentation results, the ratio of the number of pixels in the enhanced region to the total number of pixels in the effective region is calculated, yielding the enhanced coverage, with a calculation accuracy better than ±2%. The coverage segmentation branch network implements... Figure 3 The function of the coverage detection branch is to output the area ratio of the enhanced region through DeepLabV3+ segmentation network, decoder sampling and boundary refinement.

[0044] The coverage segmentation branch network includes an encoder and a decoder. The encoder takes process-sensitive features (formed by channel concatenation of enhanced process category probabilities and fused feature maps) as input and generates high-level semantic features using a dilated spatial pyramid pooling module. The decoder takes these high-level semantic features as input, upsamples them via transposed convolution, and outputs the enhanced region segmentation result. Based on this, it outputs the enhanced region coverage rate. The enhanced region segmentation result includes pixel-level background, enhanced regions, and unenhanced regions. The enhanced region coverage rate is equal to the ratio of the number of pixels in the enhanced regions to the total number of pixels in the total effective regions. The total effective regions refer to the areas outside the background.

[0045] The enhanced region coverage can be calculated as follows: take the enhanced region category probability map from the segmented Softmax probability map obtained by the coverage segmentation branch network, perform threshold binarization and morphological opening operation to remove noise, and then calculate the ratio of the number of mask pixels to the total number of effective region pixels.

[0046] When the backbone network is a shared feature extraction backbone network based on EfficientNet-B4, a preprocessed three-channel fluorescence image is input, and four levels of feature maps are extracted sequentially. The feature maps at each level are fused across layers through a feature pyramid network to construct a multi-scale feature hierarchy from P3 to P6, ultimately outputting a P3 fused feature map, which is simultaneously fed into subsequent classification and segmentation branches. Based on the P3 feature map, the process discrimination branch network uses an attention mechanism and dual-pooling fusion for classification, while the coverage segmentation branch network uses an ASPP module and upsampling for segmentation; the outputs process type probability and coverage value.

[0047] In another exemplary embodiment of this application, the method further includes an offline training phase, involving data augmentation, transfer learning strategies, and model solidification. The training data covers three scenarios: shot peening, laser shock peening, and no peening. The training process of the deep learning dual-branch network can then be replaced by steps 201-205.

[0048] Step 201: Collect fluorescence images of the surfaces of multiple samples; the sample surfaces are covered by three types of scenarios: different strengthening processes and no strengthening.

[0049] Step 202: Perform data enhancement on the acquired fluorescence image to obtain a data-enhanced fluorescence image; the data enhancement includes elastic transformation, color jitter, and random noise injection.

[0050] Data augmentation can effectively alleviate class imbalance and improve model robustness.

[0051] Step 203: Perform scene annotation and segmentation region masking on each data-enhanced fluorescence image to obtain the training dataset; the scene annotation includes enhancement process category annotation and non-enhanced annotation; the segmentation region includes background, enhanced region and non-enhanced region.

[0052] Professional annotation tools are used to annotate the images with process categories and reinforced area masking.

[0053] Step 204: Set the loss function of the process discrimination branch network to the label smoothing cross-entropy loss function, and the loss function of the coverage segmentation branch network to a weighted loss function composed of cross-entropy loss, Dice loss, and boundary Focal loss. The expression of the weighted loss function is as follows: ; in , , These are the weighting coefficients for each loss. Let cross-entropy be the loss function. The Dice loss function, The boundary Focal loss function.

[0054] The smoothing factor in the label-smoothed cross-entropy loss function can be set to 0.1 to suppress overfitting. Class weights are introduced into the cross-entropy loss to balance the proportions of the three classes: shot peening, laser shock peening, and no peening. The weighted loss function improves the coverage calculation accuracy to better than ±2%. The expression for the label-smoothed cross-entropy loss function is: ; in, The label smoothing cross-entropy loss function; An index for the original, true labels; One-hot encoding of the original real label; For the model to predict the first The softmax probability of the class; This is a smoothing factor. For category weights.

[0055] Step 205: Based on the training dataset, train the deep learning dual-branch network using the label smooth cross-entropy loss function and the weighted loss function to obtain the trained deep learning dual-branch network.

[0056] During training, the training dataset can be divided into a training set, a test set, and a validation set. In the model solidification stage, the optimal model weights are selected by the performance metrics of the validation set, serialized into a lightweight inference format, and stored to complete the offline training stage task.

[0057] The classification accuracy of the process discrimination branch network is no less than 98%.

[0058] Furthermore, after system startup, it receives raw fluorescence images in real time and loads offline-fixed model weights from local storage to complete the inference environment initialization. Standardization preprocessing is then performed on the input images.

[0059] For example, the process discrimination branch network employs a transfer learning strategy: before training the deep learning dual-branch network, the process discrimination branch network is obtained through pre-training. The training process through transfer learning is as follows: first, pre-training is performed on the ImageNet dataset, followed by fine-tuning on a constructed fluorescent surface image dataset, using the FocalLoss loss function to handle the class imbalance problem.

[0060] To address the issues of subjective nature in manual inspection and poor adaptability of traditional machine vision, this application presents a method that establishes a "chemical enhancement—optical acquisition—intelligent analysis" inspection workflow, integrating three major functional modules: surface pretreatment, perception, and analysis. The pretreatment module enhances the visibility of surface microstructures using fluorescence infiltration technology; the perception module acquires images using a high-resolution optical system; and the analysis module, based on deep learning algorithms, constructs a dual-branch network consisting of a process discrimination branch network and a coverage segmentation branch network, introducing an attention mechanism to improve feature focusing capabilities. The innovations of this application include fluorescence signal enhancement, multi-task network collaborative optimization, and process perception architecture design.

[0061] The overall process of this application method is as follows: Figure 3 As shown, a surface strengthening process detection method with wide applicability was constructed by deeply integrating fluorescent penetrant technology and deep learning algorithms. Furthermore, a neural network architecture was designed to accurately distinguish between shot peening and laser shock peening processes. Experiments show that the process classification accuracy exceeds 98%, and the coverage calculation accuracy is better than ±2%, making it suitable for efficient and accurate detection of critical components such as aero-engine blades.

[0062] Compared with the prior art, the beneficial effects of the method in this application are as follows: In terms of detection sensitivity, this application innovatively introduces a fluorescent penetration process to transform the random pits on the shot-peened surface and the micro-pits on the laser-shock-strengthened surface into high-contrast images, thereby increasing the visual discrimination of different strengthening processes by orders of magnitude and breaking through the technical bottleneck of traditional visible light imaging, which is limited by feature size, reflectivity, and material differences.

[0063] In terms of process identification algorithms, this application innovatively applies deep learning algorithms to the intelligent identification and differentiation of enhanced surface features. It employs a deep classification architecture based on EfficientNet-B3 or EfficientNet-B4 backbone networks and convolutional block attention mechanisms, combined with a dual weighting mechanism of channel and spatial attention. This allows the grid to automatically focus on the differences in the distribution patterns of fluorescent points, accurately capturing the array patterns of two different enhancement processes on the workpiece surface. Simultaneously, it utilizes transfer learning strategies and a label smoothing loss function to achieve high-precision discrimination of three different feature surfaces under limited labeled sample conditions. Using deep learning algorithms, the classification accuracy can reach over 96%, constructing a highly reliable intelligent decision-making system.

[0064] Regarding the coverage detection algorithm, this application adopts a DeepLabV3+ encoder-decoder structure combined with a hole space pyramid pooling module to capture multi-scale contextual information in parallel. Through joint optimization of Dice loss and boundary Focal loss, it directly enhances the measurement of regional overlap and the localization of sample boundaries, which significantly improves the pixel-level extraction accuracy of enhanced regional boundaries. The coverage calculation accuracy is better than ±2%, meeting the industrial needs of automated coverage detection.

[0065] Based on the same inventive concept, this application also provides a surface strengthening process identification and coverage detection system for implementing the surface strengthening process identification and coverage detection method described above. The solution provided by this system is similar to the implementation scheme described in the above method. Therefore, the specific limitations of one or more surface strengthening process identification and coverage detection system embodiments provided below can be found in the limitations of the surface strengthening process identification and coverage detection method described above, and will not be repeated here.

[0066] In one exemplary embodiment, such as Figure 4 As shown, a surface strengthening process identification and coverage detection system is provided, including a sensing device and a computer 11.

[0067] The sensing device is used to acquire fluorescence images of the workpiece surface after fluorescence treatment. The computer 11 is used to output the enhancement process category probability and enhancement region coverage of the workpiece surface based on the fluorescence image of the workpiece surface using a trained deep learning dual-branch network. The deep learning dual-branch network includes: a feature extraction module, a process discrimination branch network, and a coverage segmentation branch network. The feature extraction module is used to input the fluorescence image of the workpiece surface and generate a fused feature map through an EfficientNet backbone network and a feature pyramid network. The process discrimination branch network uses a convolutional block attention mechanism to obtain the enhancement process category probability from the fused feature map. The coverage segmentation branch network uses an encoder-decoder architecture semantic segmentation network to determine the enhancement region segmentation result based on the enhancement process category probability and the fused feature map, and then determines the enhancement region coverage based on the enhancement region segmentation result.

[0068] As an optional implementation, the sensing device includes: an optical support 1, a light source support 7, a moving platform 8, an industrial camera 2, an imaging lens 3, and an ultraviolet excitation light source 6. The optical support 1 is a cantilever structure, with the industrial camera 2 mounted at the end of the cantilever 9; the imaging lens 3 is integrated below the industrial camera 2. The light source support 7 is mounted on the optical support 1 and located below the imaging lens 3; the ultraviolet excitation light source 6 is fixed on the light source support 7 and is used to excite fluorescence on the surface of the workpiece after fluorescence treatment. The moving platform 8 is located below the light source support 7 and is used to place the workpiece and move it. The signal output terminal of the industrial camera 2 is connected to the signal input terminal of the computer 11; the industrial camera 2 is used to acquire fluorescence images of the workpiece surface after fluorescence excitation through the imaging lens 3 and transmit them to the computer 11.

[0069] In this implementation, the optical support 1 is a cantilever structure with high rigidity and vibration resistance. The support has a height of 500mm in the Z direction and a cantilever 9 with a length of 100mm. This ensures the geometric stability of the image during visual acquisition. The support is equipped with a precision moving platform 8, which enables accurate positioning and scanning of the workpiece in the XY plane, with a positioning accuracy better than ±10μm and a repeatability better than ±5μm.

[0070] Industrial Camera 2 uses a CMOS high-sensitivity camera with a resolution of no less than 5 million pixels and a dynamic range of no less than 12 bits to capture fluorescence signals.

[0071] Imaging lens 3 is an FA lens with magnification adjustable from 2x to 5x, a depth of field covering the typical depth of surface-enhanced pits, and a telecentricity better than 0.1. ° This is to eliminate perspective distortion and ensure measurement accuracy.

[0072] The ultraviolet excitation source 6 is a 365nm wavelength LED ring light source, i.e., a high-power LED array with a wavelength of 365nm, and the irradiance is not less than 10mW / cm² at a working distance of 100mm. 2 It is equipped with a narrow-band filter to suppress visible light components. The light source layout adopts a multi-angle ring structure and is arranged coaxially with the lens to ensure the uniformity of fluorescence excitation.

[0073] like Figure 4 The hardware configuration and layout shown depict an optical support 1 with an aluminum alloy cantilever structure. A camera clamp 4 is mounted at the end of the cantilever 9, allowing the vision system to move in the Z-direction via a focusing slider 5. Furthermore, the camera clamp 4 integrates an industrial camera 2 and an imaging lens 3 (a high-precision fixed-focus lens) with a magnification of 3x, a depth of field of 80μm, and a telecentricity of 0.05. °This effectively eliminates perspective distortion caused by the curved surface of the blades. The camera is a 5-megapixel CMOS camera with a 12-bit dynamic range, equipped with a 365nm narrowband filter to suppress ambient light interference. Furthermore, a light source bracket 7 is installed below the industrial camera 2 and the imaging lens 3. A ring-shaped ultraviolet excitation light source 6 is fixed on the light source bracket 7, with an irradiance of 15mW / cm² at a working distance of 100mm. 2 The light source support 7 is coaxially arranged with the imaging lens 3. Furthermore, an XY two-dimensional precision moving platform is fixed on the stage 10 of the optical support 1, with a stroke of 100mm × 100mm, a positioning accuracy of ±8μm, and a repeatability of ±3μm. This enables the movement of the titanium alloy plate sample.

[0074] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0075] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for identifying surface strengthening processes and detecting coverage, characterized in that, include: Fluorescent treatment is applied to the surface of the workpiece; Acquire fluorescence images of the workpiece surface after fluorescence treatment; Based on the fluorescence image of the workpiece surface, a trained deep learning dual-branch network is used to output the probability of the strengthening process category and the coverage of the strengthening area of ​​the workpiece surface. The deep learning dual-branch network includes: a feature extraction module, a process discrimination branch network, and a coverage segmentation branch network; The feature extraction module is used to input the fluorescence image of the workpiece surface and generate a fused feature map through the EfficientNet backbone network and the feature pyramid network; the process discrimination branch network adopts the convolutional block attention mechanism to obtain the enhanced process category probability from the fused feature map; the coverage segmentation branch network adopts the semantic segmentation network with encoder-decoder architecture, determines the enhanced region segmentation result based on the enhanced process category probability and the fused feature map, and then determines the enhanced region coverage based on the enhanced region segmentation result.

2. The surface strengthening process identification and coverage detection method according to claim 1, characterized in that, Fluorescent treatment of the workpiece surface includes: The workpiece surface is cleaned using organic solvents; Apply the fluorescent penetrant to the cleaned workpiece surface and allow it to penetrate into the pits on the workpiece surface. The residual fluorescent penetrant on the surface of the workpiece is removed by water washing to obtain a cleaned workpiece; Dry the cleaned workpieces. A white developer is applied to the surface of the dried workpiece using a spraying method, which adsorbs the fluorescent penetrant onto the surface layer of the workpiece.

3. The surface strengthening process identification and coverage detection method according to claim 1, characterized in that, Acquire fluorescence images of the workpiece surface after fluorescence treatment, including: Real-time reception of raw fluorescence images of the workpiece surface after fluorescence treatment; The original fluorescence image is preprocessed to obtain a fluorescence image of the workpiece surface after fluorescence processing; the preprocessing includes: nonlocal mean filtering for noise reduction, adaptive histogram equalization for contrast enhancement, geometric correction for distortion elimination, and image normalization.

4. The surface strengthening process identification and coverage detection method according to claim 1, characterized in that, The feature extraction module includes: a cascaded EfficientNet backbone network and a feature pyramid network; The EfficientNet backbone network is used to extract multi-scale features from the fluorescence image of the workpiece surface; Feature pyramid networks are used to fuse multi-scale features to generate fused feature maps.

5. The surface intensification process identification and coverage detection method of claim 1, wherein, The process discrimination branch network includes: a convolutional block attention module, a dual pooling layer, a Dropout layer, and a fully connected layer connected in sequence; The convolutional block attention module is used to perform dual attention enhancement on the fused feature map to obtain the enhanced feature map; The dual pooling layer is used to perform global average pooling and max pooling in parallel on the enhanced feature map and then concatenate them into a global feature vector. After overfitting is suppressed by the Dropout layer, the global feature vector is then output through a fully connected layer to enhance the probability of the process category.

6. The surface intensification process identification and coverage detection method of claim 1, wherein, The coverage segmentation branch network includes: an encoder and a decoder; The encoder takes the process-sensitive features formed by channel splicing of enhanced process category probabilities and fused feature maps as input, and uses the void space pyramid pooling module to generate high-level semantic features. The decoder is used to input the high-level semantic features, output the enhanced region segmentation result through transposed convolution upsampling, and then output the enhanced region coverage rate based on the enhanced region segmentation result; the enhanced region segmentation result includes pixel-level background, enhanced region and unenhanced region; the enhanced region coverage rate is equal to the ratio of the number of pixels in the enhanced region to the total number of pixels in the effective region; wherein, the total effective region refers to the region other than the background.

7. The surface strengthening process identification and coverage detection method according to claim 1, characterized in that, The training process of the deep learning dual-branch network includes: Fluorescence images of the surfaces of multiple samples were acquired; the sample surfaces were covered by three different strengthening processes and unstrengthened surfaces. The acquired fluorescence image is subjected to data augmentation to obtain a data-enhanced fluorescence image; the data augmentation includes elastic transformation, color dithering, and random noise injection; Each data-enhanced fluorescence image is labeled with scene annotation and segmentation region mask marking to obtain a training dataset; the scene annotation includes enhancement process category annotation and no enhancement annotation; the segmentation region includes background, enhanced region and no enhancement region; The loss function of the process discrimination branch network is set to the label smoothing cross-entropy loss function, and the loss function of the coverage segmentation branch network is a weighted loss function composed of cross-entropy loss, Dice loss and boundary Focal loss. Based on the training dataset, the deep learning dual-branch network is trained using the label smooth cross-entropy loss function and the weighted loss function to obtain the trained deep learning dual-branch network.

8. The surface enhancement process identification and coverage detection method of claim 1 or 7, wherein, The process discrimination branch network adopts a transfer learning strategy: the process discrimination branch network is obtained through pre-training before training the deep learning dual-branch network.

9. A surface enhancement process identification and coverage detection system, comprising: include: Sensing devices and computers; The sensing device is used to acquire fluorescence images of the workpiece surface after fluorescence treatment; The computer is used to output the probability of the strengthening process category and the coverage of the strengthening area of ​​the workpiece surface based on the fluorescence image of the workpiece surface using a trained deep learning dual-branch network. The deep learning dual-branch network includes: a feature extraction module, a process discrimination branch network, and a coverage segmentation branch network; The feature extraction module is used to input the fluorescence image of the workpiece surface and generate a fused feature map through the EfficientNet backbone network and the feature pyramid network; the process discrimination branch network adopts the convolutional block attention mechanism to obtain the enhanced process category probability from the fused feature map; the coverage segmentation branch network adopts the semantic segmentation network with encoder-decoder architecture, determines the enhanced region segmentation result based on the enhanced process category probability and the fused feature map, and then determines the enhanced region coverage based on the enhanced region segmentation result.

10. The surface intensification process recognition and coverage detection system of claim 9, wherein, The sensing device includes: an optical bracket, a light source bracket, a mobile platform, an industrial camera, an imaging lens, and an ultraviolet excitation light source; The optical support is a cantilever structure, with an industrial camera mounted at the end of the cantilever; the imaging lens is integrated below the industrial camera. The light source support is mounted on the optical support and located below the imaging lens; an ultraviolet excitation light source is fixed on the light source support, which is used to excite the fluorescence on the surface of the workpiece after fluorescence treatment; The moving platform is located below the light source bracket. The moving platform is used to place the workpiece and move the workpiece. The signal output terminal of the industrial camera is connected to the signal input terminal of the computer; the industrial camera is used to acquire fluorescence images of the workpiece surface after fluorescence excitation through the imaging lens and transmit them to the computer.

Images