Tire defect detection method and device and computer readable storage medium

Abstract

The invention discloses a tire defect detection method and device and a computer readable storage medium, and relates to the technical field of tire defect detection.The method comprises the steps that image information of a tire is collected based on visual inspection, and surface defect information of the tire is recognized according to the image information; acquiring a tire reflection wave when the high-frequency sound wave penetrates through the tire body of the tire, and identifying internal structure defect information of the tire according to the tire reflection wave; acquiring an eddy current signal of the tire based on an electromagnetic induction technology, and identifying surface or near-surface defect information of a tire bead and a cord fabric layer of the tire according to the eddy current signal; and determining a defect detection result of the tire according to the surface defect information, the internal structure defect information and the surface or near-surface defect information of the tire bead and the cord fabric of the tire. Therefore, the method combines the visual detection result, the reflected wave recognition result and the logistics change recognition result to complete tire defect detection, and the detection comprehensiveness and accuracy are improved.

Description

Technical Field

[0001] This application relates to the field of tire defect detection technology, and in particular to a tire defect detection method, detection device and computer-readable storage medium. Background Technology

[0002] In the tire manufacturing and usage process, defect detection of tires (especially key components such as the tread, sidewall, and carcass ply) is a core element in ensuring driving safety and product durability. As the only part of a car that comes into contact with the ground, even minor defects in the tire's interior or surface (such as bubbles, cracks, or broken cords) can lead to serious safety accidents like tire blowouts at high speeds. Therefore, accurate defect detection is crucial.

[0003] In existing technologies, defects are identified by detecting changes in eddy currents in the conductive materials inside the tire (such as the metal cords in the tire carcass). However, this method has certain limitations and reduces the accuracy of detection. Summary of the Invention

[0004] This application aims to at least partially address one of the technical problems in the related art. Therefore, the first objective of this application is to propose a tire defect detection method that combines visual inspection results, reflected wave recognition results, and material flow change recognition results to complete tire defect detection, thereby improving the comprehensiveness and accuracy of the detection.

[0005] The second objective of this application is to provide a tire defect detection device.

[0006] The third objective of this application is to provide a computer-readable storage medium.

[0007] To achieve the above objectives, a first aspect of this application proposes a tire defect detection method, which includes: acquiring tire image information based on visual detection and identifying tire surface defect information based on the image information; acquiring tire reflected waves when high-frequency sound waves penetrate the tire carcass and identifying tire internal structural defect information based on the tire reflected waves; acquiring tire eddy current signals based on electromagnetic induction technology and identifying surface or near-surface defect information of the tire bead and ply layers based on the eddy current signals; and determining the tire defect detection result based on the surface defect information, internal structural defect information, and surface or near-surface defect information of the tire bead and ply layers.

[0008] According to the tire defect detection method of this application, firstly, image information of the tire is acquired based on visual inspection, and surface defect information of the tire is identified based on the image information. Then, the tire reflected wave is obtained when high-frequency sound waves penetrate the tire carcass, and internal structural defect information of the tire is identified based on the reflected wave. Additionally, eddy current signals of the tire are acquired based on electromagnetic induction technology, and surface or near-surface defect information of the tire's bead and ply layers is identified based on the eddy current signals. Finally, the tire defect detection result is determined by combining the surface defect information, internal structural defect information, and surface or near-surface defect information of the tire's bead and ply layers. Therefore, this method combines visual inspection results, reflected wave identification results, and material flow change identification results to complete tire defect detection, improving the comprehensiveness and accuracy of the detection.

[0009] In addition, the tire defect detection method according to the above embodiments of this application may also have the following additional technical features: According to one embodiment of this application, identifying surface defect information of a tire based on image information includes: preprocessing the image information to obtain a target image; inputting the target image into a trained improved ResNet model to output a defect classification result and a corresponding localization result; and determining surface defect information based on the defect classification result and the corresponding localization result.

[0010] According to one embodiment of this application, the trained improved ResNet model includes a backbone network, an adaptive multimodal feature enhancement module, a classification branch, and a contrastive learning branch. The target image is input into the trained improved ResNet model to output defect classification results and corresponding localization results. This includes: inputting the target image into the backbone network to output a multi-scale feature map, wherein the backbone network includes a first residual block, a second residual block, a third residual block, and a fourth residual block, and the multi-scale feature map is generated based on the feature maps output by the first residual block, the second residual block, the third residual block, and the fourth residual block, respectively; inputting the multi-scale feature map and camera metadata into the adaptive multimodal feature enhancement module to output an enhanced feature map, wherein the adaptive multimodal feature enhancement module includes a spectral correction layer and a spatial-channel joint attention module; and inputting the enhanced feature map into the classification branch to output defect classification results and corresponding localization results.

[0011] According to one embodiment of this application, the tire defect detection method further includes: acquiring defect-free automobile tire images from a production line historical image library, and constructing a defect-free sample library based on the defect-free automobile tire images and corresponding tire types; constructing a defect parameter library based on preset defect types and corresponding defect type parameter ranges; constructing a tire parameter library based on preset tire material parameter types and corresponding material parameter ranges; performing random defect injection and tire characteristic adaptation adjustment on defect-free samples in the defect-free sample library based on the defect parameter library and the tire parameter library to generate multiple synthetic defect images, wherein the multiple synthetic defect images include corresponding data annotations; performing data augmentation on the multiple synthetic defect images to obtain multiple synthetic data; acquiring multiple real data, and dividing the multiple synthetic data and the multiple real data according to preset ratios to obtain a first training set and a first test set corresponding to the multiple synthetic data, and a second training set and a second test set corresponding to the multiple real data; merging the first training set and the second training set to obtain a mixed training set, and merging the first test set and the second test set to obtain a mixed test set, and training and testing an improved ResNet model based on the mixed training set and the mixed test set to obtain a trained improved ResNet model.

[0012] According to one embodiment of this application, identifying surface or near-surface defect information of the tire bead and ply layer based on eddy current signals includes: decomposing the eddy current signal into multiple time-frequency atoms based on synchronous compressed wavelet transform; identifying noise atoms among the multiple time-frequency atoms and determining a wavelet denoising threshold based on the energy distribution of the noise atoms; performing denoising filtering on the eddy current signal based on the wavelet denoising threshold to obtain a denoised time-domain signal; performing motion artifact compensation on the denoised time-domain signal to obtain a preprocessed time-domain signal; performing time-domain feature extraction, frequency-domain feature extraction, and defect morphology reconstruction feature extraction on the preprocessed time-domain signal to obtain a time-domain feature vector, a frequency-domain feature vector, defect morphology parameters, and corresponding uncertainties; determining a target feature vector based on the time-domain feature vector, frequency-domain feature vector, defect morphology parameters, and corresponding uncertainties; and classifying defects based on the target feature vector using a physical constraint neural network to identify surface or near-surface defect information of the tire bead and ply layer.

[0013] According to one embodiment of this application, defect morphology reconstruction features are extracted from the preprocessed time-domain signal to obtain defect morphology parameters and corresponding uncertainties. This includes: establishing a nonlinear mapping relationship between eddy current signals and defect morphology based on Maxwell's equations, and constructing an objective function based on the nonlinear mapping relationship; and performing a minimization iterative process on the objective function using the Levenberg-Marquardt algorithm based on the preprocessed time-domain signal to obtain the defect morphology parameters and their uncertainties.

[0014] According to one embodiment of this application, a method for classifying defects in a target feature vector based on a physically constrained neural network to identify surface or near-surface defect information of the tire's bead and ply layers includes: embedding the discrete form of Maxwell's equations into the physically constrained neural network as a regularization term of the loss function to construct a target loss function; optimizing and constraining the target feature vector based on the target loss function to obtain a physically constrained feature vector; inputting the physically constrained feature vector into a CNN (Convolutional Neural Network) model to output a defect type probability; calibrating the confidence of the defect type probability based on Platt scaling to obtain a calibrated defect type probability; and generating tire reflection waves based on the calibrated defect type probability to identify internal structural defect information of the tire.

[0015] According to one embodiment of this application, the tire defect detection method further includes: acquiring historical inspection records, eddy current detection results, and visual inspection results, and defining multiple nodes based on the historical inspection records, eddy current detection results, and visual inspection results; calculating the causal relationship strength between the multiple nodes through Granger causality test to obtain the corresponding edge weights; constructing a spatiotemporal causal graph based on the multiple nodes and their corresponding edge weights; dynamically adjusting the fusion weights between the actual eddy current detection results and the actual visual inspection results based on the spatiotemporal causal graph, and determining the joint confidence level based on the adjusted fusion weights; and triggering a defect tracing process when the joint confidence level exceeds 0.95, so as to infer the root cause of the defect through a Bayesian network.

[0016] To achieve the above objectives, a second aspect of this application provides a tire defect detection device, comprising: a first identification module for acquiring image information of the tire based on visual detection and identifying surface defect information of the tire based on the image information; a second identification module for acquiring tire reflected waves when high-frequency sound waves penetrate the tire carcass and identifying internal structural defect information of the tire based on the tire reflected waves; a third identification module for acquiring eddy current signals of the tire based on electromagnetic induction technology and identifying surface or near-surface defect information of the tire's bead and ply layers based on the eddy current signals; and a detection output module for determining the tire defect detection result based on the surface defect information, internal structural defect information, and surface or near-surface defect information of the tire's bead and ply layers.

[0017] The tire defect detection device according to an embodiment of this application acquires tire image information based on visual inspection using a first identification module, and identifies surface defect information of the tire based on the image information. A second identification module acquires tire reflected waves when high-frequency sound waves penetrate the tire carcass, and identifies internal structural defect information of the tire based on the reflected waves. A third identification module acquires eddy current signals of the tire based on electromagnetic induction technology, and identifies surface or near-surface defect information of the tire's bead and ply layers based on the eddy current signals. A detection output module determines the tire defect detection result based on the surface defect information, internal structural defect information, and surface or near-surface defect information of the tire's bead and ply layers. Therefore, this device combines visual inspection results, reflected wave identification results, and material flow change identification results to complete tire defect detection, improving the comprehensiveness and accuracy of the detection.

[0018] To achieve the above objectives, a third aspect of this application provides a computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the steps of the tire defect detection method proposed in the first aspect.

[0019] According to the computer-readable storage medium of the present application embodiment, when the stored computer program is executed by a processor, the above-described tire defect detection method is implemented, and the comprehensiveness and accuracy of the detection are improved based on the above-described tire defect detection method.

[0020] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0021] Figure 1 This is a flowchart of a tire defect detection method according to an embodiment of this application; Figure 2 This is a connection diagram of a tire defect detection device according to an embodiment of this application. Detailed Implementation

[0022] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0023] The tire defect detection method, detection device, and computer-readable storage medium proposed in this application are described below with reference to the accompanying drawings.

[0024] Figure 1 This is a flowchart of a tire defect detection method according to an embodiment of this application.

[0025] like Figure 1 As shown, the tire defect detection method of this application embodiment may include the following steps: S1: Collects tire image information based on visual inspection and identifies tire surface defect information based on the image information.

[0026] In other words, visual inspection technology is used to acquire images or three-dimensional data of car tires using industrial cameras or 3D (Three-Dimensional) vision systems. Image processing algorithms (such as edge detection and texture analysis) and deep learning models are then used to identify surface defects (such as tread cracks, sidewall bulges, and scratches) and dimensional deviations to obtain surface defect information.

[0027] S2, acquire the tire reflection wave when high-frequency sound waves penetrate the tire carcass, and identify internal structural defect information of the tire based on the tire reflection wave.

[0028] For example, high-frequency sound waves (2-5MHz) can be used to penetrate the tire carcass, and the internal structure can be analyzed through reflected waves to detect defects such as carcass delamination, cord inclusions, or air bubbles, thereby identifying information about internal structural defects in the tire.

[0029] S3 acquires eddy current signals from the tire based on electromagnetic induction technology, and identifies surface or near-surface defect information of the tire's bead and ply layer based on the eddy current signals.

[0030] In other words, by using the principle of electromagnetic induction, surface or near-surface defects (such as broken or corroded steel cords) of the tire bead and ply are identified through changes in eddy currents, so as to obtain information on surface or near-surface defects of the tire bead and ply.

[0031] S4. Determine the tire defect detection results based on surface defect information, internal structural defect information, and surface or near-surface defect information of the tire's bead and ply layers.

[0032] The final tire defect detection result is generated by combining the surface defect information, internal structural defect information, and surface or near-surface defect information of the bead and ply identified by the above three methods. For example, the surface defect information, internal structural defect information, and surface or near-surface defect information of the bead and ply can be integrated into a defect detection table, or the surface defect information, internal structural defect information, and surface or near-surface defect information of the bead and ply can be combined for mutual verification, and the final defect detection result is generated according to the defect probability. There are no specific restrictions.

[0033] This method combines visual inspection results, reflected wave recognition results, and logistics change recognition results to complete tire defect detection, improving the comprehensiveness and accuracy of the detection.

[0034] In one embodiment of this application, identifying surface defect information of a tire based on image information includes: preprocessing the image information to obtain a target image; inputting the target image into a trained improved ResNet model to output defect classification results and corresponding localization results; and determining surface defect information based on the defect classification results and corresponding localization results.

[0035] Specifically, visual inspection technology is used to acquire images or 3D data of automobile tires using industrial cameras or 3D vision systems. Image processing algorithms (such as edge detection and texture analysis) and deep learning models are then used to identify surface defects (such as tread cracks, sidewall bulges, and scratches) and dimensional deviations. This includes the following steps: Image information is obtained by photographing the surface of car tires using an industrial camera; Image information is preprocessed, such as denoising, enhancement, and correction, to obtain the preprocessed image, which is the target image. The target image is input into a trained, improved ResNet model. The model performs defect detection, outputting defect classification and localization results. The classification results include defect type probabilities (tread cracks, sidewall bulges, scratches), and the localization results include a defect heatmap (224×224, value range [0,1]), specifically the defect location coordinates (x,y), defect size (length × width, unit: pixels), and confidence score (0~1). Thus, the trained, improved ResNet model outputs defect classification results (defect type probabilities (tread cracks, sidewall bulges, scratches)) and localization results (defect heatmap (224×224, value range [0,1])), providing surface defect information.

[0036] In some embodiments of this application, the trained improved ResNet model includes a backbone network, an adaptive multimodal feature enhancement module, a classification branch, and a contrastive learning branch. Inputting a target image into the trained improved ResNet model to output defect classification results and corresponding localization results includes: inputting the target image into the backbone network to output a multi-scale feature map, wherein the backbone network includes a first residual block, a second residual block, a third residual block, and a fourth residual block, and the multi-scale feature map is generated based on the feature maps output by the first residual block, the second residual block, the third residual block, and the fourth residual block, respectively; inputting the multi-scale feature map and camera metadata into the adaptive multimodal feature enhancement module to output an enhanced feature map, wherein the adaptive multimodal feature enhancement module includes a spectral correction layer and a spatial-channel joint attention module; and inputting the enhanced feature map into the classification branch to output defect classification results and corresponding localization results.

[0037] Specifically, the improved ResNet model after training includes a backbone network, an adaptive multimodal feature enhancement module, a classification branch, and a contrastive learning branch (in this embodiment, the backbone network is shared between the training and actual use phases). The following provides a detailed explanation of the processing and recognition process of the target image using the improved ResNet model after training.

[0038] First, the target image output values ​​are processed by a backbone network. Feature extraction is performed through the backbone network to obtain a multi-scale feature map. The backbone network consists of a first residual block, a second residual block, a third residual block, and a fourth residual block set sequentially.

[0039] Taking a target image of 224×224×3 as an example, the 224×224×3 image is first input into the first residual block to output a 64×112×112 feature map (F1). The processing of the first residual block involves using 64 convolutional kernels of size 7×7 with a stride of 2. Then, F1 is input into the second residual block to output a 64×56×56 feature map (F2). The processing of the second residual block involves using 64 convolutional kernels of size 3×3 with a stride of 1. Next, F2 is input into the third residual block to output a 128×28×28 feature map (F3). The processing of the third residual block involves using 128 convolutional kernels of size 3×3 with a stride of 2. Finally, F3 is input into the fourth residual block to output a 256×14×14 feature map. The feature map (F4) is generated by processing the fourth residual block using 256 3×3 convolutional kernels with a stride of 2. Multi-scale feature maps are extracted from the outputs of the first, second, and third residual blocks: F1 (64×112×112, low-level detail features), F2 (64×56×56, mid-level edge features), F3 (128×28×28, high-level semantic features), and F4 to form the multi-scale feature map.

[0040] Then, the multi-scale feature maps and camera metadata (such as exposure time t and gain g) are input into the adaptive multimodal feature enhancement module to obtain the enhanced feature maps. The adaptive multimodal feature enhancement module includes a spectral correction layer and a spatial-channel joint attention module.

[0041] Specifically, the F1 (64×112×112, low-level detail features) and camera metadata are input to the spectral correction layer to output the illumination-normalized feature map F1_corr (64×112×112). The spectral correction layer processes the data as follows: the brightness of F1 is adjusted using learnable parameters to simulate different exposure conditions. F 1,corrected = F1⊙(1+α·Δt)+β·Δg; Among them, F 1,corrected The corrected feature map is F1_corr; F1 is the input feature map (from the output of the first residual block of ResNet-18, with a size of 64×112×112); ⊙ represents element-wise multiplication (Hadamard product); α is the exposure time calibration coefficient (unit: ms). 1) Controlling the impact of exposure time variations on feature map brightness; Δt is the exposure time increment (unit: ms), representing the difference between the current exposure time and the reference exposure time; β is the gain calibration coefficient (unit: dB). 1) The degree of influence of control gain changes on the brightness of the feature map; Δg is the gain increment (unit: dB), representing the difference between the current gain and the reference gain. Where α = 0.05 / ms and β = 0.1 / dB are calibration coefficients.

[0042] F1_corr, F2, and F3 are then input into the spatial-channel joint attention module to output enhanced feature maps F_enhanced1 (64×112×112), F_enhanced2 (64×56×56), and F_enhanced3 (128×28×28). The specific processing procedure of the spatial-channel joint attention module is as follows: By applying a 3×3 depthwise separable convolution to F1_corr through the spatial attention branch, a spatial weight map W_space1 (112×112×1) is generated; a 5×5 dilated convolution (rate=2) is applied to F2 to generate W_space2 (56×56×1), which is then upsampled to 112×112; after applying 7×7 global pooling to F3, a 1×1 convolution is applied to generate W_space3 (1×1×1), which is then broadcast to 112×112.

[0043] The spatial weights are fused based on the following formula: Wspace_final=Sigmoid(W_space1+Upsample(W_space2)+Broadcast(W_space3)); Global average pooling is performed on F1_corr using a channel attention branch to generate a 64-dimensional vector. Channel weights W_channel1 (64×1) are then generated using a fully connected layer (64→16→64). A similar operation is performed on F2 to generate W_channel2 (64×1). After global average pooling on F3, W_channel3 (128×1) is generated and downsampled to 64 dimensions.

[0044] Channel weights are fused based on the following formula: Wchannel_final=Softmax(W_channel1+W_channel2+W_channel3).

[0045] The final weights are obtained through dynamic weighting: Wfinal = Wspace_final ⊙ Wchannel_final F1_corr, F2, and F3 are weighted according to the following formula: F_enhanced=(F1⊙Wfinal)⊕(F2⊙Wfinal)⊕(F3⊙Wfinal) Here, ⊕ represents the feature splicing operation.

[0046] The outputs are the enhanced feature maps F_enhanced1 (64×112×112), F_enhanced2 (64×56×56), and F_enhanced3 (128×28×28).

[0047] Finally, the enhanced feature map F_enhanced3 (128×28×28) is input into the classification branch to identify the probability of the defect type (tread crack, sidewall bulge, scratch) and the corresponding localization result. The processing procedure of the classification branch is: global average pooling → fully connected layer (256→3) → Softmax activation.

[0048] In addition, the improved ResNet model also includes a contrastive learning branch, where the input to the contrastive learning branch is the original F3 (128×28×28) and F_enhanced3 (128×28×28), which are processed by global average pooling → fully connected layer (256→128) → L2 normalization respectively. The loss function is NT-Xent loss (Normalized Temperature-scaled Cross Entropy Loss), which forces the original features and the enhanced features to be close to each other in the hyperspherical space.

[0049] According to one embodiment of this application, the tire defect detection method further includes: acquiring defect-free automobile tire images from a production line historical image library, and constructing a defect-free sample library based on the defect-free automobile tire images and corresponding tire types; constructing a defect parameter library based on preset defect types and corresponding defect type parameter ranges; constructing a tire parameter library based on preset tire material parameter types and corresponding material parameter ranges; performing random defect injection and tire characteristic adaptation adjustment on defect-free samples in the defect-free sample library based on the defect parameter library and the tire parameter library to generate multiple synthetic defect images, wherein the multiple synthetic defect images include corresponding data annotations; performing data augmentation on the multiple synthetic defect images to obtain multiple synthetic data; acquiring multiple real data, and dividing the multiple synthetic data and the multiple real data according to preset ratios to obtain a first training set and a first test set corresponding to the multiple synthetic data, and a second training set and a second test set corresponding to the multiple real data; merging the first training set and the second training set to obtain a mixed training set, and merging the first test set and the second test set to obtain a mixed test set, and training and testing an improved ResNet model based on the mixed training set and the mixed test set to obtain a trained improved ResNet model.

[0050] Specifically, before the improved ResNet model is applied to defect detection, it needs to undergo a training process to ensure that the detection results meet the requirements of the scenario. The test set and test case set used to train the above model can be obtained as follows: First, obtain defect-free car tire images from the production line history image library. Select clear, defect-free samples covering different tire types (such as radial tires and bias tires) from the defect-free car tire images and unify their size to 224×224 pixels, normalizing them to the range of [0,1] to construct a defect-free sample library (for example, the defect-free sample library includes 1000 defect-free sample images covering 3 tire types).

[0051] Simultaneously, a defect parameter library is constructed based on preset defect types, such as tread cracks, sidewall bulges, and exposed cords, and the corresponding defect type parameter ranges. Taking the preset defect types including tread cracks, sidewall bulges, and exposed cords as an example, the defect type parameter ranges may include: tread cracks: length l∈[5, 50] pixels, width w∈[1, 5] pixels, depth d∈[0.1, 2.0] mm; sidewall bulges: diameter d∈[5, 30] mm, position (x,y)∈[0,224]×[0,224]; exposed cords: length l∈[3, 20] pixels, width w∈[1, 3] pixels. Thus, multiple sets of random parameter combinations are randomly generated to construct the defect parameter library (for example, this defect parameter library may contain 1000 sets of random parameter combinations).

[0052] A tire parameter library is constructed based on preset tire material parameter types (e.g., tire tread depth, tread hardness, cord material) and their corresponding parameter ranges. Taking preset tire material parameter types including tire tread depth, tread hardness, and cord material as an example, the parameter ranges are: Tread depth: 3mm, 5mm, 7mm (corresponding to different wear levels); Tread hardness: 60 Shore A, 65 Shore A, 70 Shore A (different rubber compounds); Cord material: steel wire, aramid fiber composite metal wire, nylon. Then, multiple random material combinations are generated to construct the tire parameter library; for example, the tire parameter library contains 9 material combinations.

[0053] Then, synthetic data is generated based on the defect-free sample library, defect parameter library, and tire parameter library. Specifically, a set of parameters (e.g., tread crack length l=10 pixels) is randomly selected from the defect parameter library, and a defect is injected at a random position (x,y) on the defect-free sample. A combination of parameters (e.g., tread depth 5mm, tread hardness 65 Shore A, cord material steel wire) is randomly selected from the tire parameter library. The image texture is adjusted according to the tread depth (texture mapping is applied to simulate the shadow effect of treads at different depths), and the defect edge morphology is adjusted according to the tread hardness (higher hardness results in sharper crack edges). The exposure time Δt∈[-10, +10]ms and gain Δg∈[-2, +2]dB are randomly selected for illumination simulation. A spectral correction layer is applied to adjust the brightness of the defect area (e.g., when Δt=+10ms, the brightness is increased by 50%). Finally, the defect location (x,y), type (tread crack / sidewall bulge / exposed cord), tire parameters (tread depth, tread hardness, cord material) and other annotation information are automatically generated to output a synthetic defect image (I_synthetic) with precise annotations.

[0054] Then, data augmentation is performed on the synthetic defect image, and the augmented synthetic data (I_synthetic_aug) is output. For example, the synthetic defect image can be processed as follows: random rotation (±10°), contrast adjustment (±20%), and noise injection (Gaussian noise σ=0.05).

[0055] The synthetic data (I_synthetic_aug) and real data (I_real) are divided into training and testing sets to construct a hybrid training set and a hybrid testing set for model pre-training. Specifically, the synthetic data is divided into training and testing sets, for example, 1000 synthetic data images are divided into a training set (800 images) and a testing set (200 images) in an 8:2 ratio; the real data is also divided into training and testing sets, for example, 500 real data images are divided into a training set (400 images) and a testing set (100 images) in an 8:2 ratio; then, the synthetic training set (800 images) and the real training set (400 images) are merged to form a hybrid training set (1200 images), and the synthetic test set (200 images) and the real test set (100 images) are merged to form a hybrid test set (300 images), thus obtaining a hybrid training set (1200 images, including labels) and a hybrid test set (300 images, including labels).

[0056] The above-described method for generating training sets can address the issues of insufficient rare defect samples and performance degradation caused by tire type variations in tire detection. Specifically, it offers the following advantages: 1. Controllable Defect Generation: Specifically, the following defects can be generated through physical model constraints: tread cracks (generating irregular paths along a specified direction, simulating the cross-effect of cracks and patterns by combining tread texture features), sidewall bulges (simulating the bulge's protruding shape through 3D spherical projection, considering edge blurring caused by rubber elasticity), and exposed cords (simulating metallic luster or fiber texture based on cord material, generating realistic visual effects by combining light reflection models).

[0057] 2. Data-driven optimization: By adjusting the diversity of synthetic data through learnable parameters (such as crack length and bulge diameter), edge cases (such as micro-cracks and micro-bulges) that are difficult to obtain by traditional methods can be covered.

[0058] 3. Adaptive Tire Characteristics: Specific implementation methods may include tread depth simulation and material characteristic modeling. Tread depth simulation involves applying texture mapping technology to generate different texture patterns based on tread depth parameters, ensuring that the synthesized image matches the real tire tread pattern. Material characteristic modeling involves adjusting the optical properties of defects based on tread hardness and cord material, so that the synthesized data conforms to the physical properties of different tires.

[0059] In addition, the training process of this improved ResNet model differs from the data used in actual applications as follows: 1. Backbone network: Improved version of ResNet-18 (consistent with the training phase).

[0060] 2. The processing flow of the adaptive multimodal feature enhancement module is the same as that of the training phase, but only the spectral correction layer of the F1 branch is enabled (because no comparative learning is required in the actual use phase).

[0061] 3. Classification and positioning head.

[0062] The input to the classification head is F_enhanced3 (128×28×28), and the structure is global average pooling → fully connected layer (256→3) → Softmax activation. The output is the probability of the defect type (tread crack, sidewall bulge, scratch).

[0063] The input to the positioning head is F4 (256×14×14), the structure is a 1×1 convolutional layer (256→1) → Sigmoid activation → upsampling to 224×224, and the output is a defect heatmap (224×224, value range [0,1]).

[0064] Compared to related techniques that only use single-scale features (such as the output of the last convolutional layer), this approach extracts multi-scale feature maps (F1-F4) through four residual blocks of an improved ResNet, covering global information from low-level details to high-level semantics. Furthermore, a spectral correction layer is introduced to compensate for illumination variations, and a spatial-channel joint attention mechanism dynamically adjusts feature weights, enabling the model to focus on defective regions.

[0065] According to one embodiment of this application, identifying surface or near-surface defect information of the tire bead and ply layer based on eddy current signals includes: decomposing the eddy current signal into multiple time-frequency atoms based on synchronous compressed wavelet transform; identifying noise atoms among the multiple time-frequency atoms and determining a wavelet denoising threshold based on the energy distribution of the noise atoms; performing denoising filtering on the eddy current signal based on the wavelet denoising threshold to obtain a denoised time-domain signal; performing motion artifact compensation on the denoised time-domain signal to obtain a preprocessed time-domain signal; performing time-domain feature extraction, frequency-domain feature extraction, and defect morphology reconstruction feature extraction on the preprocessed time-domain signal to obtain a time-domain feature vector, a frequency-domain feature vector, defect morphology parameters, and corresponding uncertainties; determining a target feature vector based on the time-domain feature vector, frequency-domain feature vector, defect morphology parameters, and corresponding uncertainties; and classifying defects based on the target feature vector using a physical constraint neural network to identify surface or near-surface defect information of the tire bead and ply layer.

[0066] According to one embodiment of this application, defect morphology reconstruction features are extracted from the preprocessed time-domain signal to obtain defect morphology parameters and corresponding uncertainties. This includes: establishing a nonlinear mapping relationship between eddy current signals and defect morphology based on Maxwell's equations, and constructing an objective function based on the nonlinear mapping relationship; and performing a minimization iterative process on the objective function using the Levenberg-Marquardt algorithm based on the preprocessed time-domain signal to obtain the defect morphology parameters and their uncertainties.

[0067] Specifically, first, the original eddy current signal (time-domain waveform, including non-stationary noise and probe jitter noise) is acquired. Then, the following steps are performed on the acquired eddy current signal in sequence to obtain the defect morphology parameters and their uncertainties.

[0068] Non-stationary noise dynamic modeling is employed: Synchrosqueezed Wavelet Transform (SWT) is applied to decompose the original eddy current signal into time-frequency atoms. The entropy value (E=) is calculated for each time-frequency atom. Atoms with an entropy value higher than the threshold (E_th=3.5), defined as Σplog p, are considered noise. The wavelet threshold is dynamically adjusted based on the energy distribution of the noise atoms. Where σ is the noise standard deviation, N is the signal length, and μ is the noise energy mean, the dynamic threshold (λ) is obtained, which is the wavelet denoising threshold, and used for subsequent wavelet denoising.

[0069] Wavelet thresholding denoising is performed: Soft thresholding is applied to the wavelet coefficients after SWT decomposition based on the wavelet denoising threshold to obtain the denoised eddy current signal. Additionally, the power frequency can be estimated through real-time Fourier analysis, and the center frequency of the notch filter is dynamically adjusted. A dual second-order notch filter is applied for filtering to obtain the denoised time-domain signal (power frequency suppression >30dB).

[0070] Motion artifact compensation is performed on the denoised time-domain signal: The denoised time-domain signal and accelerometer data (sampling rate 1kHz) are input into the acceleration-signal coupling model, which establishes the relationship between probe acceleration (a) and signal phase shift (Δ). The nonlinear model (Δ) =k•a2, where k is the material correlation coefficient and a is the probe acceleration (measured by an accelerometer). Phase correction is achieved through extended Kalman filter (EKF), thereby outputting the phase-corrected time-domain signal as the preprocessed time-domain signal.

[0071] Feature extraction is performed on the preprocessed time-domain signal. The feature extraction process includes the following sub-processes: 1. Temporal feature extraction: Calculate the peak value of the preprocessed time-domain signal mean ,variance The time-domain feature vector (3D) is output based on the results of equality, mean, and variance.

[0072] 2. Frequency domain feature extraction: For the preprocessed time-domain signal, apply the Short-Time Fourier Transform (STFT) (window length 256, overlap 50%) to calculate the dominant frequency and band energy (0-1kHz) of the preprocessed time-domain signal, and output the frequency domain feature vector (2D) based on the dominant frequency and band energy.

[0073] 3. Extracting and reconstructing features of defect morphology: A regularized inversion algorithm is applied to the preprocessed time-domain signal: Based on Maxwell's equations, a nonlinear mapping relationship is established between the eddy current signal and the defect morphology (fracture length l of the steel cord, corrosion area s); Iterative optimization: The objective function is minimized using the Levenberg-Marquardt algorithm. ),in, This is the measured signal. The signal is the prediction signal, α is the regularization coefficient, and R(l,s) is the sparse constraint term for the defect morphology. The output is the defect morphology parameters (l,s) and their uncertainty. .

[0074] Therefore, the target feature vector (3-dimensional time domain + 2-dimensional frequency domain + 4-dimensional morphology parameters = 9-dimensional) is determined based on the time domain feature vector, frequency domain feature vector, defect morphology parameters, and corresponding uncertainties.

[0075] Defect classification based on target feature vector: Defect classification is performed on the target feature vector (9 dimensions) based on a physical constraint neural network to output surface or near-surface defect information of the tire bead and ply.

[0076] In this embodiment, eddy current detection is based on the principle of electromagnetic induction. An alternating magnetic field induces eddy currents in the conductive materials (such as steel cords) of the tire bead and ply layer. Defects (such as broken or corroded steel cords) alter the eddy current path, causing a change in the impedance of the detection coil. By analyzing the impedance change signal, quantitative screening of surface and near-surface defects in the tire bead and ply layer can be achieved.

[0077] According to one embodiment of this application, a physical constraint neural network is used to classify defects in target feature vectors to identify surface or near-surface defect information of the tire's bead and ply layers. This includes: embedding the discrete form of Maxwell's equations into the physical constraint neural network as a regularization term of the loss function to construct a target loss function; optimizing and constraining the target feature vector based on the target loss function to obtain a physically constrained feature vector; inputting the physically constrained feature vector into a CNN model to output defect type probabilities; calibrating the confidence of the defect type probabilities based on Platt scaling to obtain calibrated defect type probabilities; and generating tire reflection waves based on the calibrated defect type probabilities to identify internal structural defect information of the tire.

[0078] Specifically, first, the target feature vector is input into the Physical Constraint Layer (PCL) to output a physically constrained feature vector. The processing includes: Electromagnetic field regularization: the discrete form of Maxwell's equations ( ,in, The curl of the magnetic field intensity is defined by Maxwell's equations, and J is the current density. The time derivative of the electric displacement vector is embedded in the neural network as a regularization term of the loss function.

[0079] Construct the loss function: L CE For cross-entropy loss, L Physics For physical constraint loss ( ), where λ=0.1 is the weighting coefficient.

[0080] Through iterative optimization, the physically constrained feature vector is output.

[0081] Then, lightweight CNN inference is performed on the physically constrained feature vectors. Specifically, the physically constrained feature vectors are input into the CNN model (input layer 64×64, convolutional layer 3×3@16, pooling layer 2×2, fully connected layer 256→3) to output the defect type probability (crack, hole, no defect), where the defect type probability is 3-dimensional.

[0082] The confidence level of the defect type probability is calibrated. Specifically, Platt scaling (T=0.8) is applied to map the original probability p to the calibrated probability p′=σ(log(p) / T), where σ is the Sigmoid function, to output the calibrated defect type probability (3D). Based on the calibrated defect type probability, tire reflection waves are generated to identify internal structural defects in the tire, such as "tread crack: 0.97".

[0083] According to one embodiment of this application, the tire defect detection method further includes: acquiring historical inspection records, eddy current detection results, and visual inspection results, and defining multiple nodes based on the historical inspection records, eddy current detection results, and visual inspection results; calculating the causal relationship strength between the multiple nodes through Granger causality test to obtain the corresponding edge weights; constructing a spatiotemporal causal graph based on the multiple nodes and their corresponding edge weights; dynamically adjusting the fusion weights between the actual eddy current detection results and the actual visual inspection results based on the spatiotemporal causal graph, and determining the joint confidence level based on the adjusted fusion weights; and triggering a defect tracing process when the joint confidence level exceeds 0.95, so as to infer the root cause of the defect through a Bayesian network.

[0084] In other words, the eddy current detection results and visual inspection results are fused. The eddy current detection results are based on eddy current change recognition and may include defect type, location, morphological parameters, and confidence level. The visual inspection results are based on visual inspection recognition output and may include defect type, location, morphological parameters, and confidence level, which can be obtained from the MES (Manufacturing Execution System). The specific processing procedure is as follows: Spatiotemporal causal graph construction: Nodes are defined in the spatiotemporal causal graph based on eddy current detection results (E), visual detection results (V), and historical detection records (H); the strength of causal relationships (GEV, GVH) between nodes is calculated through Granger causality test to define edge weights in the spatiotemporal causal graph, thus obtaining the spatiotemporal causal graph (including nodes and edge weights).

[0085] Dynamic weight allocation: Confidence propagation is performed, that is, the fusion weights are dynamically adjusted based on the strength of causal relationships in the spatiotemporal causal graph. Perform joint decision-making, i.e., determine the joint confidence level. .

[0086] Defect tracing: If the joint confidence score is >0.95, the defect tracing process is triggered, and the root cause of the defect (rubber material defect / vulcanization process problem / cord braiding defect) is inferred through Bayesian Network (BN) to output the tracing conclusion (1-dimensional).

[0087] This allows for the combination of eddy current testing results and visual testing results to output joint testing results (including defect type, location, overall confidence level, and source tracing conclusion).

[0088] In summary, the application of non-stationary noise dynamic modeling in this embodiment has the following advantages: 1. Improved signal-to-noise ratio: Through synchronous compressed wavelet transform (SWT) and entropy noise estimation, the SNR is improved, significantly improving signal quality under complex working conditions and enabling more accurate identification of minor defects in tire cords.

[0089] 2. Adaptive threshold: Dynamically adjust the wavelet threshold (λ) to avoid oversmoothing or undersmoothing, and retain more tire defect features, such as subtle steel cord rust marks.

[0090] 3. Strong performance: Suitable for non-steady noise environments (such as mechanical vibration in tire production lines and power fluctuations in testing equipment).

[0091] 4. High computational efficiency: SWT's time-frequency focusing reduces computational load, and its real-time performance meets the high-speed inspection requirements of tire production lines.

[0092] The application of defect morphology reconstruction features has the following advantages: 1. Quantitative characterization: Through regularized inversion algorithm, the quantitative measurement of tire defects (such as the broken length of steel cord and the area of ​​rust) is realized, reducing errors and providing accurate data for tire quality assessment.

[0093] 2. Defect source tracing: Correlation analysis between morphological parameters and process parameters revealed implicit relationships such as "vulcanization temperature deviation > 5℃ → sidewall bulge defect rate increases by 15%".

[0094] 3. Reduce missed detections: Morphology reconstruction can detect minute defects that are difficult to identify by traditional methods (such as microcracks in the tread with a depth of 0.02 mm).

[0095] 4. Process optimization: Provides data support for adjusting tire production parameters, such as optimizing weaving tension based on cord breakage to reduce defect rate.

[0096] The application of physically constrained neural networks has the following advantages: 1. Physical consistency: By embedding the Maxwell equations into the loss function, the model output conforms to the laws of electromagnetic fields, reducing false predictions of tire cord defects.

[0097] 2. Improved generalization ability: Improved accuracy in cross-tire type testing (e.g., from radial tires to bias-ply tires).

[0098] 3. High data efficiency: Reduces reliance on labeled data, reduces the amount of training samples required, and lowers the difficulty of collecting tire defect samples.

[0099] 4. High interpretability: Physical constraints make the model decision-making process transparent, making it easier for engineers to adjust tire testing parameters.

[0100] The application of spatiotemporal causal reasoning has the following advantages: 1. Multimodal fusion: By dynamically allocating weights through Granger causality testing, the joint confidence level is improved, thereby enhancing the reliability of tire defect detection.

[0101] 2. Defect tracing: The root cause of defects is inferred based on Bayesian networks, which improves the accuracy of tracing and facilitates the improvement of tire production processes.

[0102] 3. Reduce false alarms: Spatiotemporal causal graphs filter out accidental interference, such as false alarms caused by temporary stains on the tire surface, thereby reducing the false alarm rate.

[0103] 4. Closed-loop feedback: The conclusions of the source tracing can directly drive the adjustment of production line parameters, such as adjusting the formula for rubber material defects, forming a closed-loop management of tire quality.

[0104] Therefore, the tire defect detection method provided in this application can achieve the following beneficial effects: 1. Technical characteristics of combining multiple detection methods.

[0105] Complementary inspection: Eddy current inspection is good at detecting surface and near-surface defects in conductive materials such as tire bead and ply, visual inspection can capture tire surface morphology information, and data fusion can comprehensively utilize information from multiple sources to improve the comprehensiveness and accuracy of inspection.

[0106] Improved detection rate: By complementing multimodal information, defects that are difficult to detect with a single detection method can be detected, such as loose cords on the tire subsurface and tiny tread cracks.

[0107] Reduce false alarm rate: Cross-validation between different detection methods can reduce false alarms, such as avoiding misjudging stains on the tire surface as defects and improving the reliability of detection.

[0108] Compared with existing technologies, traditional methods rely on a single detection method, which makes it difficult to fully cover all types of tire defects. In contrast, this application achieves more comprehensive defect screening through multimodal detection.

[0109] 2. Technical features for dynamic modeling of non-stationary noise.

[0110] Significantly improved signal-to-noise ratio: Through synchronous compressed wavelet transform and entropy noise estimation, the SNR is improved, effectively suppressing noise under complex working conditions in tire production lines, such as mechanical vibration and power fluctuations.

[0111] Adaptive threshold adjustment: Dynamically calculates wavelet thresholds to avoid oversmoothing or undersmoothing, preserving more tire defect features, such as subtle steel cord rust marks.

[0112] Compared with existing technologies: traditional methods use fixed thresholds, which cannot adapt to changes in noise during tire testing, resulting in missed detections or false alarms.

[0113] 3. Technical features for reconstructing defect morphology.

[0114] Quantitative characterization of defects: Based on the regularized inversion algorithm of Maxwell's equations, quantitative measurement of tire defects (such as the broken length of steel cord and the depth of tread cracks) is achieved, reducing errors.

[0115] Defect tracing capability: Correlation analysis between morphological parameters and process parameters revealed implicit relationships such as "vulcanization temperature deviation > 5℃ → sidewall bulge defect rate increases by 15%", providing data support for process optimization.

[0116] Compared with existing technologies, traditional methods can only determine the existence of tire defects, but cannot provide morphological information, making it difficult to guide process improvement.

[0117] 4. Technical characteristics of physical constraint neural networks.

[0118] Enhanced physical consistency: By embedding the Maxwell equations into the loss function, the model output conforms to the laws of electromagnetic fields, reducing false predictions of tire cord defects.

[0119] Improved generalization ability: Increases accuracy and reduces reliance on labeled data in cross-tire type testing (e.g., from radial tires to bias-ply tires).

[0120] Compared with existing technologies, traditional deep learning models lack physical constraints and are prone to failure in scenarios involving different tire types.

[0121] 5. Regarding the characteristics of spatiotemporal causal reasoning.

[0122] Multimodal fusion optimization: Weights are dynamically allocated through Granger causality tests, thereby improving joint confidence.

[0123] Precise Defect Source Tracing: Bayesian networks infer the root cause of defects, improving the accuracy of source tracing and providing support for the closed-loop system of tire quality.

[0124] Compared with existing technologies, traditional fusion methods use simple weighting and do not consider the causal relationship between tire test data, resulting in unreliable fusion results.

[0125] 6. Technical characteristics of adaptive feedback control.

[0126] Real-time production line response: The PID controller reduces the delay in adjusting tire production parameters (such as vulcanization temperature and cord weaving tension) to meet the production line cycle time requirements.

[0127] The defect rate decreased significantly: the response time for systemic defects (such as three consecutive defects of the same type) was shortened, resulting in a lower defect rate.

[0128] Compared to existing technologies, traditional methods rely on manual intervention, have a slow response speed, and are difficult to adapt to the dynamic needs of tire production lines.

[0129] In summary, the tire defect detection method according to the embodiments of this application firstly acquires tire image information based on visual inspection, and identifies surface defect information of the tire based on the image information. It then acquires the tire reflection wave when high-frequency sound waves penetrate the tire carcass, and identifies internal structural defect information of the tire based on the tire reflection wave. Additionally, it acquires eddy current signals of the tire based on electromagnetic induction technology, and identifies surface or near-surface defect information of the tire's bead and ply layers based on the eddy current signals. Finally, it combines the surface defect information, internal structural defect information, and surface or near-surface defect information of the tire's bead and ply layers to determine the tire defect detection result. Therefore, this method combines visual inspection results, reflection wave identification results, and material flow change identification results to complete tire defect detection, improving the comprehensiveness and accuracy of the detection.

[0130] Corresponding to the above embodiments, this application also proposes a tire defect detection device.

[0131] like Figure 2 As shown, the tire defect detection device of this application embodiment may include: a first identification module 10, a second identification module 20, a third identification module 30, and a detection output module 40.

[0132] The first identification module 10 is used to acquire tire image information based on visual detection and identify tire surface defect information based on the image information; the second identification module 20 is used to acquire tire reflected waves when high-frequency sound waves penetrate the tire carcass and identify tire internal structural defect information based on the tire reflected waves; the third identification module 30 is used to acquire tire eddy current signals based on electromagnetic induction technology and identify surface or near-surface defect information of tire bead and ply layer based on the eddy current signals; the detection output module 40 is used to determine the tire defect detection result based on surface defect information, internal structural defect information, and surface or near-surface defect information of tire bead and ply layer.

[0133] According to one embodiment of this application, the first identification module 10 identifies surface defect information of the tire based on image information, specifically for: preprocessing the image information to obtain a target image; inputting the target image into a trained improved ResNet model to output defect classification results and corresponding localization results; and determining surface defect information based on the defect classification results and corresponding localization results.

[0134] According to one embodiment of this application, the trained improved ResNet model includes a backbone network, an adaptive multimodal feature enhancement module, a classification branch, and a contrastive learning branch. The first recognition module 10 inputs the target image into the trained improved ResNet model to output defect classification results and corresponding localization results. Specifically, it is used to: input the target image into the backbone network to output a multi-scale feature map, wherein the backbone network includes a first residual block, a second residual block, a third residual block, and a fourth residual block, and the multi-scale feature map is generated based on the feature maps output by the first residual block, the second residual block, the third residual block, and the fourth residual block, respectively; input the multi-scale feature map and camera metadata into the adaptive multimodal feature enhancement module to output an enhanced feature map, wherein the adaptive multimodal feature enhancement module includes a spectral correction layer and a spatial-channel joint attention module; and input the enhanced feature map into the classification branch to output defect classification results and corresponding localization results.

[0135] According to one embodiment of this application, the first identification module 10 is further configured to: acquire defect-free automobile tire images from a production line historical image library, and construct a defect-free sample library based on the defect-free automobile tire images and corresponding tire types; construct a defect parameter library based on preset defect types and corresponding defect type parameter ranges; construct a tire parameter library based on preset tire material parameter types and corresponding material parameter ranges; perform random defect injection and tire characteristic adaptation adjustment on defect-free samples in the defect-free sample library based on the defect parameter library and tire parameter library to generate multiple synthetic defect images, wherein the multiple synthetic defect images include corresponding data annotations; perform data augmentation on the multiple synthetic defect images to obtain multiple synthetic data; acquire multiple real data, and divide the multiple synthetic data and multiple real data according to preset ratios to obtain a first training set and a first test set corresponding to the multiple synthetic data, and a second training set and a second test set corresponding to the multiple real data; merge the first training set and the second training set to obtain a mixed training set, and merge the first test set and the second test set to obtain a mixed test set, so as to train and test the improved ResNet model based on the mixed training set and the mixed test set to obtain the trained improved ResNet model.

[0136] According to one embodiment of this application, the third identification module 30 identifies surface or near-surface defect information of the tire's bead and ply layer based on eddy current signals. Specifically, it is used to: decompose the eddy current signal into multiple time-frequency atoms based on synchronous compressed wavelet transform; identify noise atoms among the multiple time-frequency atoms and determine a wavelet denoising threshold based on the energy distribution of the noise atoms; perform denoising filtering on the eddy current signal based on the wavelet denoising threshold to obtain a denoised time-domain signal; perform motion artifact compensation on the denoised time-domain signal to obtain a preprocessed time-domain signal; perform time-domain feature extraction, frequency-domain feature extraction, and defect morphology reconstruction feature extraction on the preprocessed time-domain signal to obtain time-domain feature vectors, frequency-domain feature vectors, defect morphology parameters, and corresponding uncertainties, and determine a target feature vector based on the time-domain feature vectors, frequency-domain feature vectors, defect morphology parameters, and corresponding uncertainties; and perform defect classification on the target feature vector based on a physical constraint neural network to identify surface or near-surface defect information of the tire's bead and ply layer.

[0137] According to one embodiment of this application, the third identification module 30 performs defect morphology reconstruction feature extraction on the preprocessed time-domain signal to obtain defect morphology parameters and corresponding uncertainties. Specifically, it is used to: establish a nonlinear mapping relationship between eddy current signal and defect morphology based on Maxwell's equations, and construct an objective function based on the nonlinear mapping relationship; and perform minimization iteration processing on the objective function using the Levenberg-Marquardt algorithm based on the preprocessed time-domain signal to obtain defect morphology parameters and their uncertainties.

[0138] According to one embodiment of this application, the third identification module 30 classifies the target feature vector based on a physically constrained neural network to identify surface or near-surface defect information of the tire's bead and ply layer. Specifically, it is used to: embed the discrete form of Maxwell's equations into the physically constrained neural network as a regularization term of the loss function to construct a target loss function; optimize and constrain the target feature vector based on the target loss function to obtain a physically constrained feature vector; input the physically constrained feature vector into a CNN model to output the defect type probability; calibrate the confidence of the defect type probability based on Platt scaling to obtain a calibrated defect type probability; and generate a tire reflection wave based on the calibrated defect type probability to identify the internal structural defect information of the tire.

[0139] According to one embodiment of this application, the detection output module 40 is further configured to: acquire historical detection records, eddy current detection results, and visual detection results, and define multiple nodes based on the historical detection records, eddy current detection results, and visual detection results; calculate the causal relationship strength between the multiple nodes through Granger causality test to obtain the corresponding edge weights; construct a spatiotemporal causal graph based on the multiple nodes and their corresponding edge weights; dynamically adjust the fusion weights between the actual eddy current detection results and the actual visual detection results based on the spatiotemporal causal graph, and determine the joint confidence level based on the adjusted fusion weights; and trigger a defect tracing process when the joint confidence level exceeds 0.95, so as to infer the root cause of the defect through a Bayesian network.

[0140] It should be noted that for details not disclosed in the tire defect detection device of this application embodiment, please refer to the details disclosed in the tire defect detection method of the above embodiment of this application, which will not be repeated here.

[0141] The tire defect detection device according to embodiments of this application acquires tire image information based on visual inspection using a first identification module, and identifies surface defect information of the tire based on the image information. A second identification module acquires tire reflected waves when high-frequency sound waves penetrate the tire carcass, and identifies internal structural defect information of the tire based on the reflected waves. A third identification module acquires eddy current signals of the tire based on electromagnetic induction technology, and identifies surface or near-surface defect information of the tire's bead and ply layers based on the eddy current signals. A detection output module determines the tire defect detection result based on the surface defect information, internal structural defect information, and surface or near-surface defect information of the tire's bead and ply layers. Therefore, this device combines visual inspection results, reflected wave identification results, and material flow change identification results to complete tire defect detection, improving the comprehensiveness and accuracy of the detection.

[0142] Corresponding to the above embodiments, this application also proposes a computer-readable storage medium.

[0143] The computer-readable storage medium of this application embodiment stores a computer program, wherein when the computer program is executed by a processor, it implements the steps of the above-described tire defect detection method.

[0144] According to the computer-readable storage medium of the present application embodiment, when the stored computer program is executed by a processor, the above-described tire defect detection method is implemented, and the comprehensiveness and accuracy of the detection are improved based on the above-described tire defect detection method.

[0145] It should be noted that the logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0146] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0147] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0148] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0149] In this application, unless otherwise expressly specified and limited, the terms "installation," "connection," "joining," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components, unless otherwise expressly limited. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0150] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A method for detecting tire defects, characterized in that, The method includes: Image information of the tire is acquired based on visual inspection, and surface defect information of the tire is identified based on the image information; Acquire the tire reflection wave when high-frequency sound waves penetrate the tire carcass, and identify internal structural defect information of the tire based on the tire reflection wave; The eddy current signal of the tire is obtained based on electromagnetic induction technology, and the surface or near-surface defect information of the tire bead and ply layer is identified based on the eddy current signal. The defect detection result of the tire is determined based on the surface defect information, the internal structural defect information, and the surface or near-surface defect information of the tire's bead and ply.

2. The tire defect detection method according to claim 1, characterized in that, Identifying surface defect information of the tire based on the image information includes: The image information is preprocessed to obtain the target image; The target image is input into the trained improved ResNet model to output defect classification results and corresponding localization results; The surface defect information is determined based on the defect classification results and the corresponding location results.

3. The tire defect detection method according to claim 2, characterized in that, The trained improved ResNet model includes a backbone network, an adaptive multimodal feature enhancement module, a classification branch, and a contrastive learning branch. The target image is input into the trained improved ResNet model to output defect classification results and corresponding localization results, including: The target image is input into the backbone network to output a multi-scale feature map. The backbone network includes a first residual block, a second residual block, a third residual block, and a fourth residual block. The multi-scale feature map is generated based on the feature maps output by the first residual block, the second residual block, the third residual block, and the fourth residual block, respectively. The multi-scale feature map and camera metadata are input into the adaptive multimodal feature enhancement module to output an enhanced feature map. The adaptive multimodal feature enhancement module includes a spectral correction layer and a spatial-channel joint attention module. The enhanced feature map is input into the classification branch to output the defect classification result and the corresponding localization result.

4. The tire defect detection method according to claim 3, characterized in that, The method further includes: Obtain defect-free car tire images from the production line historical image library, and construct a defect-free sample library based on the defect-free car tire images and the corresponding tire types; A defect parameter library is constructed based on preset defect types and corresponding defect type parameter ranges; A tire parameter library is constructed based on preset tire material parameter types and corresponding material parameter ranges; Based on the defect parameter library and the tire parameter library, random defect injection and tire characteristic adaptation adjustment are performed on the defect-free samples in the defect-free sample library to generate multiple synthetic defect images, wherein the multiple synthetic defect images include corresponding data annotations; Data augmentation is performed on the multiple synthetic defect images to obtain the multiple synthetic data; Multiple real data sets are acquired, and the multiple synthetic data sets and the multiple real data sets are divided according to a preset ratio to obtain a first training set and a first test set corresponding to the multiple synthetic data sets, and a second training set and a second test set corresponding to the multiple real data sets. The first training set and the second training set are combined to obtain a mixed training set, and the first test set and the second test set are combined to obtain a mixed test set. The improved ResNet model is trained and tested based on the mixed training set and the mixed test set to obtain the trained improved ResNet model.

5. The tire defect detection method according to claim 1, characterized in that, Identifying surface or near-surface defect information of the tire's bead and ply layers based on the eddy current signal includes: The eddy current signal is decomposed into multiple time-frequency atoms based on synchronous compressed wavelet transform; Identify noise atoms among the plurality of time-frequency atoms, and determine the wavelet denoising threshold based on the energy distribution of the noise atoms; The eddy current signal is denoised and filtered according to the wavelet denoising threshold to obtain the denoised time domain signal. Motion artifact compensation is performed on the denoised time-domain signal to obtain the preprocessed time-domain signal; The preprocessed time-domain signal is subjected to time-domain feature extraction, frequency-domain feature extraction, and defect morphology reconstruction feature extraction to obtain time-domain feature vector, frequency-domain feature vector, defect morphology parameters, and corresponding uncertainties. The target feature vector is determined based on the time-domain feature vector, the frequency-domain feature vector, the defect morphology parameters, and the corresponding uncertainties. Defect classification is performed on the target feature vector based on a physical constraint neural network to identify surface or near-surface defect information of the tire's bead and ply.

6. The tire defect detection method according to claim 5, characterized in that, The preprocessed time-domain signal is subjected to defect morphology reconstruction feature extraction to obtain defect morphology parameters and corresponding uncertainties, including: A nonlinear mapping relationship between eddy current signals and defect morphology is established based on Maxwell's equations, and an objective function is constructed based on the nonlinear mapping relationship. Based on the preprocessed time-domain signal, the objective function is minimized iteratively using the Levenberg-Marquardt algorithm to obtain the defect morphology parameters and their uncertainties.

7. The tire defect detection method according to claim 6, characterized in that, Defect classification is performed on the target feature vector based on a physically constrained neural network to identify surface or near-surface defect information of the tire's bead and ply, including: The discrete form of the Maxwell equations is embedded into the physical constraint neural network as a regularization term of the loss function to construct the target loss function. Based on the target loss function, the target feature vector is optimized and constrained to obtain the physically constrained feature vector; The physically constrained feature vector is input into the CNN model to output the probability of the defect type; The confidence level of the defect type probability is calibrated based on Platt scaling to obtain the calibrated defect type probability. The tire reflection wave is generated based on the calibrated defect type probability to identify the internal structural defect information of the tire.

8. The tire defect detection method according to any one of claims 1-7, characterized in that, The method further includes: Acquire historical detection records, eddy current detection results, and visual detection results, and define multiple nodes based on the historical detection records, eddy current detection results, and visual detection results; The edge weights are obtained by calculating the strength of the causal relationship between multiple nodes using the Granger causality test. Construct a spatiotemporal causal graph based on the multiple nodes and their corresponding edge weights; The fusion weights between the actual eddy current detection results and the actual visual detection results are dynamically adjusted based on the spatiotemporal causality graph, and the joint confidence is determined based on the adjusted fusion weights. If the joint confidence level exceeds 0.95, the defect tracing process is triggered to infer the root cause of the defect through a Bayesian network.

9. A tire defect detection device, characterized in that, The device includes: The first identification module is used to acquire image information of the tire based on visual detection, and to identify surface defect information of the tire based on the image information; The second identification module is used to acquire the tire reflection wave when high-frequency sound waves penetrate the tire body, and to identify the internal structural defect information of the tire based on the tire reflection wave. The third identification module is used to acquire the eddy current signal of the tire based on electromagnetic induction technology, and to identify surface or near-surface defect information of the tire's bead and ply layer based on the eddy current signal. The detection output module is used to determine the defect detection result of the tire based on the surface defect information, the internal structural defect information, and the surface or near-surface defect information of the tire's bead and ply layers.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, wherein the computer program, when executed by a processor, implements the steps of the tire defect detection method as described in any one of claims 1 to 8.

Images