A defect localization system for carbon fiber bicycle frame production

Abstract

The application discloses a kind of defect positioning systems for carbon fiber bicycle frame production, specifically relates to material nondestructive testing technical field, two-dimensional physical field image is collected;Three-dimensional convolutional neural network model trained by transfer learning is inverted into internal defect three-dimensional structure;Incremental migration module dynamically adjusts the proportion of real data and simulation data according to regulation strategy;Cognitive metric module calculates physical consistency index and cognitive uncertainty entropy in real time;Cognitive state regulation module constructs two-dimensional state space with the two, generates differentiated regulation strategy according to the area where state point is located, and guides incremental learning process;The application builds model cognitive state by real-time evaluation of physical consistency and cognitive entropy, and dynamically adjusts data ratio and physical constraint accordingly, realizes differentiated incremental learning regulation, solves the problems of lack of explainability in model prediction and poor adaptability of static learning, significantly improves the reliability, self-adaptive ability and whole-process intelligent level of defect detection.

Description

Technical Field

[0001] This invention relates to the field of nondestructive testing technology for materials, and more specifically, to a defect location system for the production of carbon fiber bicycle frames. Background Technology

[0002] Carbon fiber bicycle frames are widely used in both racing and high-end consumer bicycles due to their lightweight and high strength. The production process of carbon fiber frames involves complex processes such as multi-layer prepreg lamination, molding, and high-temperature curing, which can easily lead to internal defects such as delamination, porosity, inclusions, and fiber breakage. These defects directly affect the frame's mechanical properties and service safety. Therefore, efficient defect detection and precise location during the production process are crucial for ensuring product quality.

[0003] In existing technologies, non-destructive testing of carbon fiber composites mainly employs methods such as ultrasonic scanning, X-ray imaging, and infrared thermography. Ultrasonic scanning, by emitting sound waves and receiving reflected signals through a probe, can detect internal delamination and porosity, but its detection efficiency is low and its adaptability to complex curved surfaces is insufficient. X-ray imaging can visually display the internal structure, but its sensitivity to micron-level delamination defects is limited, and it poses radiation safety concerns. Infrared thermography infers internal defects by analyzing the evolution of the surface temperature field, offering advantages such as non-contact operation and high detection speed; however, the interpretation of thermal images relies on empirical judgment, making it difficult to accurately quantify the three-dimensional morphology of defects. These traditional methods are mostly offline inspections, presenting results as two-dimensional images or one-dimensional signals, making it difficult to visually represent the three-dimensional spatial morphology of defects.

[0004] In recent years, some researchers have attempted to apply convolutional neural networks to defect identification in carbon fiber composite materials. By training models, they have automatically extracted defect features from inspection images, improving detection efficiency. Simultaneously, the introduction of 3D reconstruction technology has made it possible to invert the 3D structure of defects from multi-view 2D images. However, existing deep learning-based defect detection systems still face the following challenges in practical applications: Existing incremental learning research mainly focuses on how to balance new and old data to avoid catastrophic forgetting. However, incremental learning strategies are mostly based on fixed presets, such as using a constant small learning rate or a fixed ratio of new to old data. These static strategies cannot be dynamically adjusted according to the model's current learning state. When the model's understanding of a certain type of defect has matured, continuing to force learning may interfere with existing knowledge; when the model has a systematic bias towards a certain type of defect, conventional incremental learning is difficult to correct quickly. Therefore, this invention proposes a defect localization system for carbon fiber bicycle frame production to solve the above problems. Summary of the Invention

[0005] To achieve the above objectives, the present invention provides the following technical solution: A defect location system for carbon fiber bicycle frame production, comprising: The image acquisition module is used to acquire two-dimensional detection images, which are the surface physical field response images of the carbon fiber bicycle frame under test under a preset excitation load. The 3D inversion module is used to process the 2D detection image based on the pre-trained 3D convolutional neural network model and output the 3D structural information of the internal defects of the carbon fiber bicycle frame to be tested. The 3D convolutional neural network model is incrementally trained from source domain simulation data to target domain real detection data through transfer learning. The incremental transfer module is used to execute the incremental learning process. During the incremental learning process, the ratio between the newly introduced real detection data and the historical simulation data used for pre-training is adjusted in real time according to the corresponding differentiated incremental learning control strategy generated by the cognitive state control module. The cognitive metrics module is used to calculate the physical consistency index and cognitive uncertainty entropy of a 3D convolutional neural network model when outputting 3D structural information in real time. The cognitive state regulation module is used to construct a two-dimensional cognitive state space composed of physical consistency index and cognitive uncertainty entropy. It maps the current cognitive state of the three-dimensional convolutional neural network model to a state point in the two-dimensional cognitive state space and generates corresponding differentiated incremental learning regulation strategies based on the cognitive state region where the state point is located.

[0006] In a preferred embodiment, the image acquisition module includes: The multi-segment heterogeneous frequency thermal excitation unit is used to apply thermal pulse excitation of different frequencies and durations according to the curvature radius and wall thickness distribution of different parts of the carbon fiber bicycle frame under test. Among them, the pulse frequency applied to the parts with smaller curvature radius and thinner wall thickness is higher and the duration is shorter. The multi-angle synchronous imaging unit includes at least three infrared thermal imagers arranged circumferentially around the carbon fiber bicycle frame under test. The imaging optical axes of the at least three infrared thermal imagers intersect at the central axis of the carbon fiber bicycle frame under test, and are used to synchronously acquire image sequences of the surface temperature field evolution of the carbon fiber bicycle frame under test at different angles during thermal pulse excitation. The layup orientation compensation unit is used to acquire the carbon fiber layup orientation information of the carbon fiber bicycle frame under test in the image acquisition area, and to perform anisotropic thermal diffusion compensation on the surface temperature field evolution image sequence based on the layup orientation information, generating a standardized thermal response image after eliminating the influence of layup orientation, which is then used as a two-dimensional detection image input to the three-dimensional inversion module.

[0007] In a preferred embodiment, the multi-segment heterogeneous frequency thermal excitation unit is specifically used for: The radius of curvature ρ and wall thickness d of different parts of the carbon fiber bicycle frame under test are obtained, and the thermal pulse excitation parameters are determined based on the radius of curvature ρ and wall thickness d. The thermal pulse excitation parameters include the pulse frequency f and the pulse duration t. The pulse frequency f satisfies f=k1 / (ρ*d)+f0 with respect to the radius of curvature ρ and the wall thickness d, and the pulse duration t satisfies t=k2*ρ*d+t0 with respect to the radius of curvature ρ and the wall thickness d. Wherein, k1 and k2 are proportionality coefficients related to the thermal diffusivity of carbon fiber composite materials, and f0 and t0 are inherent parameters of the excitation system. The smaller the radius of curvature and the thinner the wall thickness, the higher the pulse frequency and the shorter the pulse duration are applied. The pulse frequency f ranges from 0.1 Hz to 10 Hz, and the pulse duration t ranges from 1 ms to 100 ms.

[0008] In a preferred embodiment, the ply orientation compensation unit is specifically used for: The carbon fiber layup orientation angle θ of the carbon fiber bicycle frame under test in the image acquisition area is obtained. The layup orientation angle θ is defined as the angle between the carbon fiber axis and the preset reference direction. A thermal diffusion anisotropy model of carbon fiber composites was established to describe the propagation law of heat in carbon fiber composites. The thermal diffusion coefficient along the carbon fiber axis is 5 to 10 times that perpendicular to the carbon fiber axis. The isothermal surface of heat propagation is ellipsoidal, and the major axis of the ellipsoid is consistent with the carbon fiber axis. In the plane perpendicular to the fiber, the diffusion rate of heat along the warp and weft directions is faster than the diffusion rate along the 45-degree direction. Based on the anisotropic thermal diffusion model and the layup orientation angle θ, a two-dimensional point diffusion function h(x,y) is constructed. The point diffusion function h(x,y) is: ; in, , σx is positively correlated with the thermal diffusivity along the fiber axis, and σy is positively correlated with the thermal diffusivity perpendicular to the fiber axis, with the value of σx / σy ranging from 5 to 10. The image of each frame in the surface temperature field evolution image sequence is deconvolved with the point spread function h(x,y) to eliminate the spatial broadening effect of the temperature field caused by the anisotropic thermal diffusion of carbon fibers and restore the true geometric shape of the defect. The standardized thermal response image after deconvolution is output and used as the input of the two-dimensional detection image to the three-dimensional inversion module.

[0009] In a preferred embodiment, the three-dimensional convolutional neural network model is used for: Receive two-dimensional detection images from multiple angles acquired by the image acquisition module, including at least three different viewpoints distributed circumferentially around the carbon fiber bicycle frame under test; Feature extraction is performed on the two-dimensional detection image from each viewpoint to generate a two-dimensional feature map for the corresponding viewpoint. Feature extraction is achieved through convolutional layers with shared weights. Two-dimensional feature maps from multiple perspectives are mapped to three-dimensional space through projection transformation to generate an initial three-dimensional feature volume. The projection transformation is based on the response characteristics of carbon fiber composite material to a preset excitation load. The response characteristics include the correlation between the X-ray attenuation coefficient or thermal diffusivity coefficient and the material density. The X-ray attenuation coefficient is suitable for X-ray excitation load scenarios, and the thermal diffusivity coefficient is suitable for thermal pulse excitation load scenarios. The correlation is pre-calibrated by the physical properties of carbon fiber composite material. Perform 3D convolution operation on the initial 3D feature volume, fuse multi-view information and eliminate reconstruction ambiguity in the projection process to generate an enhanced 3D feature volume; The enhanced 3D feature volume is upsampled layer by layer to a preset 3D spatial resolution, and a 3D voxel representation is generated through the output layer. The value of each voxel in the 3D voxel representation represents the probability value of the existence of internal defects at that spatial location. The output is a three-dimensional voxel representation of the three-dimensional structural information of the internal defects of the carbon fiber bicycle frame under test.

[0010] In a preferred embodiment, the physical consistency index is calculated as follows: Obtain the 3D structural information of the defect predicted by the 3D convolutional neural network model, and use it as the first prediction vector; The three-dimensional structural information of the defect obtained by forward deduction based on the fracture mechanics theoretical model of carbon fiber composite material is used as the second prediction vector. The fracture mechanics theoretical model is established based on the excitation response characteristics and material property parameters in the two-dimensional detection image. The first and second prediction vectors are mapped to the complex space and expressed in complex form, where the modulus of the complex number represents the defect amplitude information and the argument of the complex number represents the defect phase information. Calculate the complex interference term between the first prediction vector and the second prediction vector. The complex interference term is defined as the inner product of the two vectors in the complex space minus half the sum of the squares of their moduli. The physical consistency index is determined based on the sign and magnitude of the complex interfering terms. When the complex interfering terms are positive, the physical consistency index is positively correlated with the magnitude of the complex interfering terms, indicating that the predictions of the physical model and the AI ​​model are in phase and reinforce each other. When the complex interfering terms are negative, the physical consistency index is negatively correlated with the absolute value of the complex interfering terms, indicating that the predictions of the physical model and the AI ​​model are in opposite phase and weaken each other. When the complex interfering terms are zero, the physical consistency index is zero, indicating that the two are orthogonal and have no interference.

[0011] In a preferred embodiment, the cognitive uncertainty entropy is calculated as follows: Obtain the prediction results of the three-dimensional convolutional neural network model for the same defect under multiple different imaging angles. The prediction results under each imaging angle include the three-dimensional structural information of the defect and its confidence distribution. The prediction results at each imaging angle are subjected to probability transformation to convert them into a probability distribution function at that angle. The probability distribution function represents the probability value of the existence of defects at each spatial location. Perform a Fourier transform on each probability distribution function to obtain its frequency domain representation, and extract the phase angle at the point of maximum amplitude in the frequency domain representation as the phase feature of the probability distribution function at that angle; Calculate the normalized cross-correlation similarity between the probability distribution functions corresponding to any two different imaging angles. The normalized cross-correlation similarity is obtained by integrating the product of the two probability distribution functions and then dividing by the product of the square roots of their respective square integrals. Complex interferometric coefficients are constructed based on the absolute value of the normalized cross-correlation similarity and the phase difference between the two imaging angles. The modulus of the complex interferometric coefficient is the absolute value of the normalized cross-correlation similarity, and the argument of the complex interferometric coefficient is the phase difference. Construct a complex interference matrix composed of complex interference coefficients. The complex interference matrix is ​​a square matrix in which the number of rows and columns are equal to the number of imaging angles. By performing eigenvalue decomposition on the complex interference matrix, multiple complex eigenvalues ​​are obtained; The complex entropy value of the complex interference matrix is ​​calculated. The real part of the complex entropy value is determined by the weighted sum of the negative logarithms after modulus normalization of each complex eigenvalue, and the imaginary part of the complex entropy value is determined by the average argument of each complex eigenvalue. The modulus of the complex entropy value is used as the cognitive uncertainty entropy.

[0012] In a preferred embodiment, the two-dimensional cognitive state space is divided into four cognitive state regions: The first cognitive region corresponds to a physical consistency index that is lower than the first preset threshold and a cognitive uncertainty entropy that is higher than the second preset threshold, indicating that the model is in a state of dual physical and cognitive imbalance. The second cognitive region corresponds to a physical consistency index that is lower than the first preset threshold and a cognitive uncertainty entropy that is not higher than the second preset threshold, indicating that the model is in a state of insufficient physical law cognition. The third cognitive region corresponds to a physical consistency index that is not lower than the first preset threshold and a cognitive uncertainty entropy that is higher than the second preset threshold, indicating that the model is in a state of insufficient data distribution cognition. The fourth cognitive region corresponds to a physical consistency index that is not lower than the first preset threshold and a cognitive uncertainty entropy that is not higher than the second preset threshold, indicating that the model is in a state of cognitive equilibrium.

[0013] In a preferred embodiment, the cognitive state regulation module is further configured to: When the state point is located in the first cognitive region, a collaborative regulation operation is performed to simultaneously enhance the weight of the physical constraint regularization term in the training process of the 3D convolutional neural network model and increase the proportion of newly introduced real detection data in incremental learning. When the state point is located in the second cognitive region, a physical orientation control operation is performed, which only increases the weight of the physical constraint regularization term and maintains the current real detection data ratio. When the state point is located in the third cognitive region, data-oriented adjustment operations are performed, which only increase the proportion of real detection data and maintain the weight of the current physical constraint regularization term. When the state point is located in the fourth cognitive region, a steady-state maintenance operation is performed to maintain the current weight of the physical constraint regularization term and the current ratio of the actual detection data unchanged.

[0014] The technical effects and advantages of this invention are as follows: This invention utilizes a cognitive metric module to calculate the physical consistency index and cognitive uncertainty entropy of a 3D convolutional neural network model in real time when outputting 3D structural information. A cognitive state control module then constructs a two-dimensional cognitive state space spanned by these two dimensions, precisely mapping the current cognitive state of the 3D convolutional neural network model to a state point within this two-dimensional cognitive state space. Differentiated incremental learning control strategies are generated based on the different cognitive state regions where each state point is located. This mechanism enables the system to perceive in real time the model's adherence to the physical laws of fracture mechanics in carbon fiber composite materials and its grasp of its own predictions. It overcomes the limitations of traditional deep learning models that only output predictions without assessing their reliability, solving the technical problems of lacking physical interpretability and difficulty in determining the reliability of predictions. When the physical consistency index is low, the system can promptly identify deviations between model predictions and physical laws; when the cognitive uncertainty entropy is high, the system can accurately determine discrepancies in predictions from multiple perspectives. Based on a comprehensive evaluation of these two dimensions, the system can fully grasp the cognitive health of the model, significantly improving the reliability and credibility of defect detection results and providing a scientific basis for subsequent control decisions.

[0015] This invention executes the incremental learning process through an incremental transfer module and dynamically adjusts the ratio between newly introduced real detection data and historical simulation data used for pre-training in real time according to the differentiated incremental learning control strategy generated by the cognitive state control module. When the state point is located in the first cognitive region of physical and cognitive imbalance, a collaborative control operation is performed to simultaneously enhance the weight of the physical constraint regularization term in the training process of the 3D convolutional neural network model and increase the ratio of newly introduced real detection data in incremental learning. When the state point is located in the second cognitive region of insufficient physical law understanding, a physical orientation control operation is performed to enhance only the weight of the physical constraint regularization term while maintaining the current ratio of real detection data unchanged. When the state point is located in the third cognitive region of insufficient data distribution understanding, a data orientation control operation is performed to increase only the ratio of real detection data while maintaining the current weight of the physical constraint regularization term unchanged. When the state point is located in the fourth cognitive region of cognitive balance, a steady-state maintenance operation is performed to maintain the current weight of the physical constraint regularization term and the current ratio of real detection data unchanged. This adaptive control mechanism completely changes the static mode of traditional incremental learning, which uses fixed data ratios and fixed learning rates. It solves the technical problems that static training models cannot adapt to fluctuations in production processes, cannot continuously optimize their own performance, and have a gradually decreasing detection accuracy when faced with changes in defect patterns. This enables the system to continuously evolve itself as the production process progresses and always maintain optimal detection performance.

[0016] This invention processes two-dimensional detection images acquired by an image acquisition module using a three-dimensional inversion module trained on a three-dimensional convolutional neural network model based on transfer learning. The output is the three-dimensional structural information of internal defects in the carbon fiber bicycle frame under test. The three-dimensional convolutional neural network model is incrementally transferred from source domain simulation data to target domain real detection data through transfer learning, effectively solving the problem of insufficient model generalization ability caused by the difference in distribution between simulation and real data. Simultaneously, a complete cognitive evaluation and learning control closed loop, consisting of a cognitive measurement module and a cognitive state control module, achieves intelligent management of the entire process from two-dimensional image acquisition to three-dimensional defect reconstruction, and from model cognitive state evaluation to dynamic adjustment of incremental learning strategies. This architecture changes the traditional defect detection system's reliance on offline detection, manual interpretation, and experience-based adjustments, solving the technical challenges of accurately quantifying the three-dimensional morphology of defects, optimizing the learning process based on the model's real-time cognitive state, and balancing adherence to physical laws with adaptation to data distribution. Through deep integration of physical models and data-driven approaches, and adaptive incremental learning driven by cognitive state, this invention provides an efficient, accurate, and adaptive defect localization solution for carbon fiber bicycle frame production, significantly improving production efficiency, reducing quality risks, and enhancing product competitiveness. Attached Figure Description

[0017] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings; Figure 1 This is a schematic diagram of a defect location system for the production of carbon fiber bicycle frames according to the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0019] Reference Figure 1 The following examples were obtained: Example 1: A defect location system for carbon fiber bicycle frame production, comprising: The image acquisition module is used to acquire two-dimensional inspection images, which are surface physical field response images of the carbon fiber bicycle frame under test under a preset excitation load. This module serves as the system's data input function. By applying the preset excitation load, it causes the internal defects of the carbon fiber bicycle frame under test to form an observable physical field response on the surface, converting the invisible internal structural information into two-dimensional image data that can be used for subsequent processing. The preset excitation load includes thermal pulse excitation or X-ray excitation, and the surface physical field response images include infrared thermograms or X-ray transmission images, providing raw data containing defect information for subsequent three-dimensional inversion.

[0020] The 3D inversion module processes 2D detection images based on a pre-trained 3D convolutional neural network model, outputting 3D structural information of internal defects in the carbon fiber bicycle frame under test. The 3D convolutional neural network model is incrementally trained from source domain simulation data to target domain real detection data through transfer learning. This module achieves a leap from 2D observation to 3D cognition. It extracts defect features from 2D detection images using the 3D convolutional neural network model, maps multi-view 2D features to 3D space through differentiable projection transformation, and then fuses multi-view information and eliminates reconstruction ambiguities through 3D convolution operations, finally outputting a voxel-level 3D defect probability distribution. The model training employs a transfer learning strategy, first pre-training on a large amount of simulation data to learn the general rules of defect response, and then incrementally transferring the model using a small amount of real detection data to adapt it to the real defect morphology in actual production.

[0021] The incremental transfer module executes the incremental learning process. During incremental learning, the ratio of newly introduced real detection data to historical simulation data used for pre-training is adjusted in real-time based on differentiated incremental learning control strategies generated by the cognitive state control module. This module is responsible for the continuous evolution of the model, constantly absorbing newly collected real defect samples during production and dynamically adjusting the mixing ratio of real data and historical simulation data. This ratio adjustment is not fixed but follows instructions from the cognitive state control module, adjusting the weight of real data accordingly when the model's cognitive state changes, ensuring the model remains in an optimal learning state.

[0022] The cognitive metrics module is used to calculate the physical consistency index and cognitive uncertainty entropy of the 3D convolutional neural network model in real time when outputting 3D structural information. This module performs a dual evaluation of the model output. The physical consistency index measures the degree of agreement between the model's predicted 3D defect structure and the theoretical model of fracture mechanics of carbon fiber composite materials, reflecting whether the prediction results conform to the physical laws of material failure. The cognitive uncertainty entropy measures the dispersion of the model's prediction results for the same defect under different imaging angles, reflecting the model's grasp of its own predictions. These two indicators characterize the model's current cognitive state from two dimensions: adherence to physical laws and grasp of data.

[0023] The cognitive state regulation module constructs a two-dimensional cognitive state space composed of a physical consistency index and cognitive uncertainty entropy. It maps the current cognitive state of the three-dimensional convolutional neural network model to a state point in this two-dimensional cognitive state space and generates corresponding differentiated incremental learning regulation strategies based on the cognitive state region where the state point is located. This module achieves a closed loop from cognitive assessment to learning regulation, constructing a two-dimensional space with the two metrics as orthogonal dimensions. The current state of the model is quantified as a point in this space. Based on the different regions the state point falls into—a region of dual physical-cognitive imbalance, a region of insufficient understanding of physical laws, a region of insufficient understanding of data distribution, or a region of cognitive equilibrium—corresponding regulation instructions are generated to direct the incremental transfer module to adjust the learning ratio or simultaneously adjust the physical constraint weights, ensuring the model consistently evolves towards a cognitive equilibrium state.

[0024] The image acquisition module includes a multi-segment heterogeneous frequency thermal excitation unit. This unit applies thermal pulses of different frequencies and durations to different parts of the carbon fiber bicycle frame under test, based on the radius of curvature and wall thickness distribution. Parts with smaller radii of curvature and thinner walls receive higher pulse frequencies and shorter durations. Carbon fiber bicycle frames have complex, irregularly shaped tubular structures with significant differences in radii of curvature and wall thickness across different parts, resulting in varying thermal conductivity and thermal response sensitivity. Using uniform thermal pulse excitation parameters cannot adequately address all areas; thin-walled areas with large curvature are prone to overheating damage, while thick-walled areas with small curvature exhibit insufficient thermal response. This unit employs a differentiated excitation strategy to match the thermal pulse parameters of each part to its geometric characteristics, ensuring a uniform and appropriate thermal response across the entire frame surface, providing a high-quality initial signal for subsequent imaging.

[0025] The multi-angle synchronous imaging unit comprises at least three infrared thermal imagers arranged circumferentially around the carbon fiber bicycle frame under test. The imaging optical axes of these three imagers intersect at the central axis of the frame, enabling the synchronous acquisition of image sequences of the surface temperature field evolution of the frame from different angles during thermal pulse excitation. The carbon fiber bicycle frame is a three-dimensional tubular structure; a single-view infrared thermal imager cannot completely record the temperature field changes across all surfaces, resulting in numerous obstructed areas and blind spots. This unit, by circumferentially arranging multiple thermal imagers and converging their optical axes at the frame's central axis, achieves full circumferential coverage without blind spots. The synchronously acquired multi-angle image sequences completely record the spatiotemporal evolution of the frame's surface temperature field after thermal pulse excitation, providing ample multi-view observation data for subsequent three-dimensional inversion.

[0026] The layup orientation compensation unit acquires the carbon fiber layup orientation information of the carbon fiber bicycle frame under test in the image acquisition area. Based on this information, it performs anisotropic thermal diffusion compensation on the surface temperature field evolution image sequence, generating a standardized thermal response image that eliminates the influence of the layup orientation. This image is then used as input to the 3D inversion module as a two-dimensional detection image. Carbon fiber composites exhibit significant anisotropic thermal diffusion characteristics; the thermal conductivity along the fiber direction is much higher than that perpendicular to the fiber. This anisotropy leads to distortion in the surface temperature field distribution, causing the thermal anomalies caused by defects to be distorted or masked by the material's inherent anisotropy. This unit acquires the layup orientation information, compensates for the original temperature field image, eliminates the influence of material anisotropy on thermal diffusion, restores the intrinsic thermal response morphology of the defects, and ensures that the final output standardized thermal response image reflects only the defect information and is independent of the layup orientation, providing clean input data for the 3D inversion module.

[0027] The multi-segment heterogeneous frequency thermal excitation unit is specifically used to: obtain the radius of curvature ρ and wall thickness d of different parts of the carbon fiber bicycle frame under test, and determine the thermal pulse excitation parameters based on the radius of curvature ρ and wall thickness d. The thermal pulse excitation parameters include the pulse frequency f and the pulse duration t. This step first measures the geometric features of the carbon fiber bicycle frame under test. The radius of curvature ρ and wall thickness d of each part of the frame are obtained through 3D scanning or design drawings. The unit of radius of curvature ρ is millimeters, and the unit of wall thickness d is millimeters. The radius of curvature reflects the degree of bending of the frame tubing, such as the small radius of curvature in the bent part of the fork and the large radius of curvature in the straight part of the top tube; the wall thickness reflects the number of carbon fiber layers, such as the large wall thickness in the joint reinforcement part and the small wall thickness in the middle of the tubing. After obtaining the geometric parameters, they are substituted into the subsequent formula to calculate the required pulse frequency f and pulse duration t for each part, realizing the personalized customization of the excitation parameters.

[0028] The pulse frequency f satisfies f = k1 / (ρ*d) + f0 with respect to the radius of curvature ρ and wall thickness d, and the pulse duration t satisfies t = k2*ρ*d + t0 with respect to the radius of curvature ρ and wall thickness d. This formula establishes a quantitative mapping relationship between the thermal pulse excitation parameters and the geometric characteristics of the vehicle frame. The formula uses a combination of inverse and direct proportions, based on the physical principle that areas with smaller radii of curvature and thinner walls have lower heat capacity and faster thermal response, requiring high-frequency short pulse excitation to generate sufficient thermal signal without overheating damage; areas with larger radii of curvature and thicker walls have higher heat capacity and slower thermal response, requiring low-frequency long pulse excitation to allow for sufficient heat penetration. These two formulas simplify the complex heat transfer problem into an engineering-calcifiable expression.

[0029] k1 and k2 are proportionality coefficients related to the thermal diffusivity of carbon fiber composites, while f0 and t0 are inherent parameters of the excitation system. The values ​​of k1 and k2 depend on the thermal diffusivity of the carbon fiber composite itself. According to existing technical literature, the measured thermal diffusivity of carbon fiber composites ranges from 0.4 to 1.6 mm² / s, fluctuating with changes in fiber volume content, porosity, and temperature. The method for determining k1 and k2 is as follows: standard values ​​of the thermal diffusivity for typical carbon fiber bicycle frame prepregs (such as T700 grade carbon fiber / epoxy resin systems) are taken and calibrated through finite element simulation. For example, when the thermal diffusivity is 1.0 mm² / s, k1 is taken as an empirical value of 5 to 20 Hz / mm², and k2 as an empirical value of 0.1 to 2 ms / mm². f0 and t0 are limiting parameters of the excitation system hardware itself, depending on the trigger delay of the infrared thermal imager and the response capability of the pulse heat source. According to existing technology, the exposure time of a short-wave infrared imaging module can be configured from 100 nanoseconds to 40 milliseconds, and the pulse width of the laser pulse generator can be below 0.5 milliseconds. Therefore, f0 is usually below 0.1 Hz (corresponding to the lowest stable output frequency of the system), and t0 is below 1 millisecond (corresponding to the shortest stable output pulse width of the system). In actual systems, f0 and t0 are determined by the specifications of the specific thermal excitation equipment (such as a flash array or laser).

[0030] The smaller the radius of curvature and the thinner the wall thickness, the higher the pulse frequency and the shorter the pulse duration applied. This qualitative rule is a direct expression of the formula. For example, in the middle region of the bicycle frame downtube, with a relatively large radius of curvature (approximately 50 mm) and moderate wall thickness (approximately 2 mm), the calculated pulse frequency is approximately 1 Hz and the pulse duration is approximately 20 milliseconds. At the junction of the head tube and top tube, with a smaller radius of curvature (approximately 15 mm) and thicker wall thickness (approximately 3 mm), the calculated pulse frequency is approximately 0.5 Hz and the pulse duration is approximately 30 milliseconds. At the bend of the rear fork, with an extremely small radius of curvature (approximately 8 mm) and thinner wall thickness (approximately 1.5 mm), the calculated pulse frequency is approximately 5 Hz and the pulse duration is approximately 8 milliseconds. This differentiated excitation ensures that the surface temperature rise of each region is basically consistent after thermal pulse excitation, avoiding overheating in thin-walled, small-curvature areas or insufficient thermal response in thick-walled, large-curvature areas.

[0031] The pulse frequency f ranges from 0.1 Hz to 10 Hz, and the pulse duration t ranges from 1 ms to 100 ms. This range is determined based on the thermal diffusion characteristics of carbon fiber composites and actual testing requirements. The lower frequency limit of 0.1 Hz corresponds to a period of 10 seconds, suitable for areas with the thickest walls and smallest curvature (such as bottom bracket joints), ensuring sufficient time for heat to penetrate inwards. The upper frequency limit of 10 Hz corresponds to a period of 0.1 seconds, suitable for areas with the thinnest walls and largest curvature (such as the tip of the rear upper fork), preventing excessive lateral heat diffusion and reduced resolution. The lower duration limit of 1 ms corresponds to an extremely short pulse, suitable for thin-walled areas to prevent heat accumulation; the upper duration limit of 100 ms corresponds to a longer pulse, suitable for thick-walled areas to ensure thermal signal strength. This range aligns with the commonly used modulation frequency range of 0.1 to 2 Hz for thermal diffusion detection of carbon fiber composites in existing technologies, and is appropriately extended according to the geometric characteristics of bicycle frames to ensure coverage of all possible application scenarios.

[0032] The layup orientation compensation unit is specifically used to: obtain the layup orientation angle θ of the carbon fiber in the image acquisition area of ​​the carbon fiber bicycle frame under test. The layup orientation angle θ is defined as the angle between the carbon fiber axis and a preset reference direction. This step first obtains the layup information of each image acquisition area on the frame surface. The carbon fiber bicycle frame is made of multiple layers of prepreg, and each layer of fiber has a specific orientation. The layup orientation angle θ is the angle between the fiber axis and a preset reference direction (such as the frame axis or the horizontal direction). The acquisition method can be based on the layup definition in the frame design drawings or measured in real time by online layup detection equipment (such as laser speckle method). For example, the downtube area of ​​the frame uses unidirectional prepreg layup, and the fiber axis is parallel to the frame axis. At this time, the layup orientation angle θ is 0 degrees. The headtube area uses cross layup to increase strength. The fiber axis of a certain layer forms a 45-degree angle with the axis. The θ of this layer is 45 degrees. This information is used to compensate for the effects of anisotropic thermal diffusion in subsequent steps.

[0033] A thermal diffusivity anisotropy model for carbon fiber composites was established to describe the heat propagation behavior within the composite material. The model shows that the thermal diffusivity along the carbon fiber axis is 5 to 10 times that perpendicular to the axis, and the isothermal surface for heat propagation is ellipsoidal, with its major axis aligned with the carbon fiber axis. Furthermore, in planes perpendicular to the fibers, the heat diffusion rate along the warp and weft directions is faster than that along the 45-degree direction. This step is based on the material's physical properties to construct a mathematical model. Carbon fiber composites exhibit anisotropic thermal conductivity due to their fiber orientation, with faster heat conduction along the fiber direction and slower conduction perpendicular to it. Existing research data shows that the thermal diffusivity along the fiber direction of typical carbon fiber epoxy composites is approximately 5 to 10 times that perpendicular to the fiber direction (the specific ratio depends on the fiber volume fraction, resin type, etc.). Therefore, the isothermal surface for heat propagation is ellipsoidal, with its major axis along the fiber direction and its minor axis perpendicular to the fiber direction. In a plane perpendicular to the fibers, due to the fabric structure (warp and weft), the diffusion rate along the warp and weft directions (i.e., the 0-degree and 90-degree directions) is faster than along the 45-degree direction.

[0034] Based on the anisotropic thermal diffusion model and the layup orientation angle θ, a two-dimensional point diffusion function h(x,y) is constructed. The point diffusion function h(x,y) is: ; in, , σx is positively correlated with the thermal diffusivity along the fiber axis, and σy is positively correlated with the thermal diffusivity perpendicular to the fiber axis, with σx / σy ranging from 5 to 10. x and y refer to the spatial coordinates within the image plane. Specifically, x and y represent the horizontal and vertical coordinates of a pixel in the image, with units of pixels. The origin of the coordinate system is usually defined at the center or upper left corner of the image. In thermal diffusivity compensation applications, h(x,y) describes the distribution of energy from an ideal point heat source located at the image origin (0,0) across surrounding pixels (x,y) after anisotropic thermal diffusivity of the carbon fiber material. Therefore, for each pixel location (x,y) in the image, the value of h(x,y) represents the proportion of heat contribution received from the origin heat source at that location. The x' and y' obtained after coordinate transformation are new coordinates obtained by rotating the original image coordinates (x,y) to align with the carbon fiber layup direction, ensuring that the major axis of the point diffusion function ellipse is aligned with the fiber axis, thus accurately simulating the directionality of anisotropic diffusion.

[0035] This step quantifies the anisotropic thermal diffusion effect into a point spread function (PSF) in image processing. The PSF describes the temperature distribution pattern formed by the thermal diffusion of an ideal point heat source on the material surface. Due to anisotropy, the PSF is a Gaussian ellipse, with its major axis determined by the layup orientation angle θ, and the ratio of its major to minor axes determined by the ratio of the thermal diffusivity coefficients (σx / σy, ranging from 5 to 10). Coordinate transformations x' and y' rotate the image coordinate system to align with the fiber direction. For example, in the region where the layup orientation angle θ = 30 degrees, the PSF is an ellipse with its major axis aligned with the 30-degree direction, and its σx (corresponding to the axial direction) is 8 times that of σy (corresponding to the vertical direction), indicating that heat diffuses more widely along the 30-degree direction. This function quantitatively describes the blurring effect of anisotropy on the thermal image.

[0036] Each frame of the surface temperature field evolution image sequence is deconvolved with the point spread function h(x,y) to eliminate the spatial broadening effect of the temperature field caused by anisotropic thermal diffusion of carbon fibers, thus restoring the true geometric shape of the defect. This step performs image restoration. The original thermal image is the convolution result of the true temperature distribution of the defect and the point spread function. The deconvolution operation solves inversely using the known point spread function to restore the unblurred thermal response image of the defect. The deconvolution algorithm can use Wiener filtering or the Richardson-Lucy iterative algorithm to deconvolve each frame of the temperature image with the point spread function, eliminating the spatial broadening and directional distortion caused by anisotropic thermal diffusion. For example, a circular layered defect, which would originally appear as an ellipse (with its major axis along the fiber direction) in anisotropic materials, is restored to a circle after deconvolution, restoring its true geometric shape.

[0037] The normalized thermal response image after deconvolution is output and used as the input to the 3D inversion module as a 2D detection image. This step outputs the processed result; the compensated image eliminates the influence of layup direction and anisotropy, and the thermal anomaly features in the image are only related to the defect itself and are independent of material orientation. These normalized images, as high-quality input data, are fed into the 3D inversion module for 3D reconstruction of the defect.

[0038] The 3D convolutional neural network model is used to: receive 2D detection images from multiple angles acquired by an image acquisition module, including at least three different viewpoints distributed circumferentially around the carbon fiber bicycle frame under test. This step inputs the multi-view normalized thermal response images or X-ray transmission images output by the image acquisition module into the neural network. For example, the image acquisition module arranges four infrared thermal imagers around the frame circumference, simultaneously acquiring data from the front, rear, left, and right directions, generating four 2D detection images with different viewpoints. Each image records the temperature field distribution on the frame surface at the same moment. These four images together constitute the network's input tensor, providing complete observation data for subsequent 3D reconstruction.

[0039] Feature extraction is performed on the 2D detection images from each viewpoint, generating corresponding 2D feature maps. Feature extraction is achieved through convolutional layers with shared weights. This step performs convolution operations independently on the image from each viewpoint to extract defect features. Shared weights mean that the four images use the same set of convolutional kernel parameters, ensuring the comparability of features extracted from different viewpoints. For example, the first layer of the network uses 64 3×3 convolutional kernels to convolve each input image, generating a 64-channel 2D feature map. High-response regions in the feature map correspond to the locations of suspected defects in the image. The shared weight design enables the network to learn viewpoint-invariant feature representations, allowing defects to be identified in the same way regardless of the viewing angle.

[0040] The two-dimensional feature maps from multiple perspectives are mapped to three-dimensional space through projection transformation to generate an initial three-dimensional feature volume. The projection transformation is based on the response characteristics of carbon fiber composites to a preset excitation load. The response characteristics include the correlation between the X-ray attenuation coefficient or thermal diffusivity coefficient and the material density. The X-ray attenuation coefficient is applicable to X-ray excitation load scenarios, while the thermal diffusivity coefficient is applicable to thermal pulse excitation load scenarios, and the correlation is pre-calibrated by the physical properties of carbon fiber composites. This step back-projects the two-dimensional feature maps into three-dimensional space to construct the initial three-dimensional feature volume. The projection transformation is not a simple geometric back-projection, but a weighted calculation based on the material's physical response characteristics. The X-ray attenuation coefficient is obtained by irradiating a standard carbon fiber specimen of known thickness with X-rays, measuring the transmission intensity, and calculating the linear attenuation coefficient according to the Beer-Lambert law. The typical X-ray attenuation coefficient of T700 carbon fiber composites is approximately 0.2 to 0.4 per centimeter. The thermal diffusivity coefficient is obtained by measuring the carbon fiber specimen using the laser flare method, recording the time required for the back side temperature to rise to half of its maximum value, and calculating the thermal diffusivity coefficient. The typical value ranges from 0.4 to 1.6 square millimeters per second. Correlation calibration refers to establishing a correspondence table between material density and response coefficient. For example, by preparing standard samples with different porosities (1%, 2%, 5%), measuring their density and response coefficient, and fitting a functional relationship, a correlation can be obtained. In practical applications, for X-ray scenes, regions with higher density have a larger X-ray attenuation coefficient, and these projection lines are given higher weight during projection transformation. For thermal pulse scenes, regions with higher density have a larger (or smaller, depending on the defect type) thermal diffusivity coefficient, and corresponding corrections are made during projection transformation. This step maps the two-dimensional feature maps (size 256×256×64) from four perspectives to a 128×128×128 three-dimensional grid through projection transformation. Each grid point accumulates projection values ​​from multiple perspectives to form an initial three-dimensional feature volume.

[0041] The initial 3D feature volume undergoes 3D convolution operations to fuse multi-view information and eliminate reconstruction ambiguity during projection, generating an enhanced 3D feature volume. This step involves 3D feature learning from the initial 3D feature volume. Although the initial 3D feature volume contains multi-view information, reconstruction ambiguity exists due to information loss and noise during projection (e.g., different defects may produce similar projections). The 3D convolution operation uses 3×3×3 convolution kernels sliding in 3D space to aggregate neighborhood voxel information and learn the 3D spatial structural features of the defect. For example, the network contains four layers of 3D convolution, each with 64 convolution kernels. As the number of layers increases, the receptive field gradually expands, enabling step-by-step modeling from local structures (e.g., small layers) to global structures (e.g., through cracks). After 3D convolution, an enhanced 3D feature volume is generated, maintaining the size of 128×128×128, but increasing the number of feature channels to 128. The feature vector of each voxel fuses multi-view information, significantly reducing ambiguity.

[0042] The enhanced 3D feature volume is upsampled layer by layer to a preset 3D spatial resolution, and a 3D voxel representation is generated through the output layer. The value of each voxel in the 3D voxel representation represents the probability of an internal defect existing at that spatial location. This step restores the feature volume to the final 3D spatial resolution and outputs the probability. Upsampling uses 3D transposed convolution, progressively reducing the feature volume size from 128... 3 Upsampled to 256 3 Simultaneously, the number of channels is gradually compressed. The output layer uses a 1×1×1 convolution to compress the feature channels to 1, and then uses a sigmoid activation function to map the output values ​​to between 0 and 1, forming the final three-dimensional voxel representation. The value of each voxel represents the probability of a defect existing at that location; for example, a voxel value of 0.9 indicates a 90% probability of a defect existing at that location, and 0.05 indicates a 5% probability. The entire three-dimensional voxel representation fully describes the three-dimensional spatial distribution of defects inside the frame, including the location, shape, size, and confidence level information of the defects.

[0043] The output is a three-dimensional voxel representation of the internal structural information of the carbon fiber bicycle frame under test. This step outputs the three-dimensional probabilistic voxels as the final result, which is then used by the subsequent cognitive measurement module to assess physical consistency and cognitive uncertainty. It can also be used for visualization or defect quantification analysis.

[0044] The physical consistency index is calculated as follows: The 3D structural information of the defect predicted by the 3D convolutional neural network model is obtained as the first prediction vector. This step receives the 3D voxel representation output by the 3D convolutional neural network model from the 3D inversion module; this representation is the first prediction vector. The first prediction vector is a 3D array whose dimensions correspond to the length, width, and height directions of the internal space of the chassis. Each element has a value between 0 and 1, representing the probability of a defect existing at that spatial location. For example, for a 256×256×256 3D voxel mesh, the first prediction vector contains approximately 16.77 million voxel values. If multiple consecutive voxel values ​​in a certain region exceed 0.8, it indicates a high probability of a defect in that region. This vector represents the defect perception learned by the AI ​​model based on multi-view image data.

[0045] The process involves acquiring three-dimensional structural information of defects obtained through forward deduction from a fracture mechanics theoretical model of carbon fiber composites. This information serves as the second prediction vector. The fracture mechanics theoretical model is established based on the excitation response characteristics and material properties in the two-dimensional detection images. This step constructs the second prediction vector based on physical principles. The fracture mechanics theoretical model is established based on the failure mechanism of carbon fiber composites, such as linear elastic fracture mechanics or the cohesive zone model, to describe the response characteristics of defects (such as delamination and cracks) under external excitation. In specific implementation, the two-dimensional detection images (such as surface temperature field distribution) acquired by the image acquisition module are used as input boundary conditions. Combined with material properties such as elastic modulus, fracture toughness, and thermal diffusivity, the model uses finite element forward solving to deduce the three-dimensional distribution of defects that may exist inside the frame under the current excitation. For example, if an abnormal temperature rise is detected on the surface of a certain area, the physical model, based on the heat conduction equation and fracture criteria, reversely calculates that this temperature rise may correspond to a delamination defect located at a depth of 2 mm and a diameter of 5 mm, and outputs the three-dimensional probability distribution of this defect as the second prediction vector. This vector represents the defect recognition based on physical laws, contrasting with the prediction of the AI ​​model.

[0046] The first and second prediction vectors are mapped to complex space, representing them as complex numbers. The modulus of the complex number represents the defect amplitude information, and the argument represents the defect phase information. This step transforms two three-dimensional vectors from the real number domain to the complex number domain for representation. The amplitude information corresponds to the probability of the defect's existence; for example, a voxel value of 0.8 is mapped to a modulus of 0.8. The phase information corresponds to the spatial positional offset characteristics of the defect, obtained through Hilbert transform or local phase calculation of the vectors, reflecting the subtle offset of the defect relative to the reference position. For example, for a defect region with a probability peak of 0.9 in the first prediction vector, its complex representation is 0.9 multiplied by exp(jφ1), where φ1 represents the small phase shift of the defect center relative to the theoretical position; the complex representation of the corresponding defect in the second prediction vector is 0.7 multiplied by exp(jφ2), where φ2 represents the spatial offset angle or morphological phase of the defect region predicted by the fracture mechanics theoretical model relative to the reference position (such as the grid origin or the theoretical defect center, i.e., the theoretical position). The introduction of complex space enables subsequent quantification of the phase consistency and amplitude consistency between the two model predictions.

[0047] The complex interference term between the first and second prediction vectors is calculated. This term is defined as the inner product of the two vectors in complex space minus half the sum of the squares of their magnitudes. This step calculates the interference effect between the two complex vectors. The inner product in complex space is defined as the sum of the complex products of corresponding positions of the two vectors. For example, for the first prediction vector A and the second prediction vector B, both are three-dimensional arrays containing 16.77 million complex numbers. The inner product is a complex number; its real part reflects the degree of overlap between the two vectors in phase, and its imaginary part reflects the relationship of orthogonal phases. Subtracting half the sum of the squares of their magnitudes eliminates the influence of their respective energies, retaining only the interaction term. This interference term quantifies the coherence between the two model predictions: when the two predictions are completely consistent, the interference term is positive and maximum; when the two predictions are completely opposite (180 degrees out of phase), the interference term is negative and maximum in absolute value.

[0048] The physical consistency index is determined based on the sign and magnitude of the complex interference term. When the complex interference term is positive, the physical consistency index is positively correlated with its magnitude, indicating that the predictions of the 3D convolutional neural network model and the fracture mechanics theory model are mutually reinforcing. When the complex interference term is negative, the physical consistency index is negatively correlated with its absolute value, indicating that the predictions of the 3D convolutional neural network model and the fracture mechanics theory model are mutually weakening. When the complex interference term is zero, the physical consistency index is zero, indicating that the two are orthogonal and have no interference. This step transforms the interference term into the final physical consistency index. The physical consistency index is a dimensionless scalar, and its value is usually normalized to between 0 and 1. For example, if the complex interfering term is calculated to be a positive value of 0.6, the physical consistency index is set to 0.6, indicating that the AI ​​model and the physical model are highly consistent and mutually corroborate each other. If the interfering term is a negative value of -0.4, the physical consistency index is set to 0.4, indicating that there is a significant deviation between the two predictions and that the AI ​​model's prediction conflicts with physical laws. If the interfering term is close to zero, the physical consistency index approaches zero, indicating that the AI ​​model's prediction has no significant correlation with physical laws. This index, as one of the outputs of the cognitive measurement module, is used for subsequent construction and control decisions of the cognitive state space.

[0049] The method for calculating cognitive uncertainty entropy is as follows: Obtain the prediction results of a 3D convolutional neural network model for the same defect from multiple different imaging angles. The prediction result for each imaging angle includes the 3D structural information of the defect and its confidence distribution. This step collects multi-view prediction data from the 3D convolutional neural network model. Assume the image acquisition module is arranged around the carbon fiber bicycle frame at four imaging angles, denoted as the first angle, second angle, third angle, and fourth angle, for example, directly in front, directly behind, left, and right of the frame. The 3D convolutional neural network model independently performs 3D inversion on the 2D detection image for each angle, outputting the 3D structural information of the defect corresponding to that angle, i.e., a 3D voxel array. The value of each voxel represents the probability of a defect existing at that location; simultaneously, it outputs the confidence score of each voxel prediction, for example, obtained through the variance of multiple forward propagations of the model. This yields four sets of prediction results, each containing a 3D array of defect probabilities and a 3D array of confidence scores, jointly describing the spatial distribution of the defect inferred from the observation of that angle and the model's confidence in this inference.

[0050] The prediction results at each imaging angle undergo a probability transformation, converting them into a probability distribution function for that angle. This probability distribution function represents the probability of a defect existing at each spatial location. This step normalizes the original probability values ​​into a valid probability distribution. For example, for the three-dimensional array of defect probabilities output from the first angle, normalization (e.g., dividing all voxel values ​​by the sum) is applied to make the sum equal to 1, thus obtaining the probability distribution function for the first angle. In this function, the value of each voxel represents the relative probability density of a defect appearing at that spatial location, and subsequent calculations are based on this distribution function. Similarly, the same transformation is performed on the outputs from the second, third, and fourth angles, resulting in four probability distribution functions.

[0051] Perform a Fourier transform on each probability distribution function to obtain its frequency domain representation. Extract the phase angle at the point of maximum amplitude in the frequency domain representation as the phase feature of the probability distribution function at that angle. This step transforms the spatial domain probability distribution to the frequency domain to extract the dominant frequency phase information. Perform a three-dimensional Fourier transform on the probability distribution function of the first angle to obtain its complex representation in the frequency domain. In this complex representation, find the frequency point with the maximum amplitude and record the corresponding phase angle, which is the phase feature of the first angle. This phase angle reflects the spatial offset characteristics of the overall defect structure projected at that angle. For example, if the defect is a circular layer perpendicular to the first angle, its dominant frequency phase may be close to zero; if the defect is tilted, the phase angle will shift. Similarly, perform the same operation on the probability distribution functions of the second, third, and fourth angles to obtain their respective phase features.

[0052] The normalized cross-correlation similarity between the probability distribution functions corresponding to any two different imaging angles is calculated. This is obtained by integrating the product of the two probability distribution functions and then dividing by the product of the square roots of their respective integrals. This step quantifies the similarity between prediction results from different angles. Taking the first and second angles as examples, the probability distribution functions of the first and second angles are multiplied together and integrated over the entire three-dimensional space to obtain the numerator. The integrals of the squares of the first and second angle probability distribution functions over the entire three-dimensional space are then calculated separately, and their square roots are multiplied together to obtain the denominator. The numerator divided by the denominator yields the normalized cross-correlation similarity between the first and second angles. This value is between -1 and +1, where +1 indicates that the two distributions are completely identical, -1 indicates that they are completely opposite, and zero indicates no correlation. By combining the four angles pairwise, a total of six similarity values ​​can be calculated.

[0053] Complex interferometric coefficients are constructed based on the absolute value of the normalized cross-correlation similarity and the phase difference between the two imaging angles. The modulus of the complex interferometric coefficient is the absolute value of the normalized cross-correlation similarity, and the argument is the phase difference. This step combines the real-valued similarity and phase difference into a complex form. For the first and second angles, the absolute value of their normalized cross-correlation similarity is taken as the modulus, and the difference between the phase features of the first and second angles is taken as the argument, thus constructing a complex interferometric coefficient. This complex number includes both the correlation in intensity between the two angle prediction results (represented by the modulus) and their consistency in spatial offset (represented by the argument). The corresponding complex interferometric coefficients are constructed similarly for all other angle pairs.

[0054] A complex interference matrix is ​​constructed, consisting of complex interference coefficients. This matrix is ​​a square matrix with both rows and columns equal to the number of imaging angles. This step arranges the complex interference coefficients for all angle pairs into a matrix form. For example, for four imaging angles, a 4x4 matrix is ​​constructed, where the element in the i-th row and j-th column is the complex interference coefficient between the i-th and j-th angles. The elements on the diagonal of the matrix are the interference coefficients between themselves, with a magnitude of one and an argument of zero, representing perfect coherence. This matrix completely records the pairwise interference relationships between all angles.

[0055] The complex interference matrix is ​​subjected to eigenvalue decomposition to obtain multiple complex eigenvalues. This step involves mathematically decomposing the complex interference matrix to solve for its eigenvalues ​​and eigenvectors. Since the matrix is ​​a complex square matrix, the eigenvalue decomposition yields the same number of complex eigenvalues ​​as the matrix dimension, with each eigenvalue containing both a real and an imaginary part. These eigenvalues ​​reflect the overall structural characteristics of the matrix, such as the intensity and phase shift of the principal components.

[0056] The complex entropy of the complex interference matrix is ​​calculated. The real part of the complex entropy is determined by the weighted sum of the negative logarithms of the normalized moduli of each complex eigenvalue, while the imaginary part is determined by the average argument of each complex eigenvalue. This step constructs the complex entropy based on the eigenvalues. First, the moduli of all complex eigenvalues ​​are summed. Then, the modulus of each eigenvalue is divided by this sum to obtain the normalized weights. The negative logarithm of each normalized weight is taken and multiplied by the weight itself. These products are then summed to obtain the real part of the complex entropy. This real part is similar to the classical Shannon entropy, measuring the consistency of predictions across angles; a smaller real part indicates greater consistency in predictions across angles. Second, the average argument (i.e., phase angle) of all complex eigenvalues ​​is calculated to obtain the imaginary part of the complex entropy. This imaginary part reflects the overall trend of phase shift. Thus, a complex entropy is obtained, with its real and imaginary parts carrying information of different dimensions.

[0057] The modulus of the complex entropy is used as the cognitive uncertainty entropy. This step takes the modulus of the complex entropy (i.e., the square root of the sum of the squares of the real and imaginary parts) as the final output cognitive uncertainty entropy. This modulus combines the level of consistency between predictions from different angles (contributed by the real part) and the magnitude of the overall phase shift (contributed by the imaginary part). When the predictions from different angles are highly consistent and there is no significant phase shift, the real part is close to zero, the imaginary part is close to zero, and the modulus is very small, indicating that the model has high cognitive certainty regarding the defects. When there is a large discrepancy in the predictions or a significant overall phase shift, the modulus increases, indicating that the model has high cognitive uncertainty. This entropy is used as the output of the cognitive measurement module to construct a two-dimensional cognitive state space and guide the dynamic adjustment of incremental learning.

[0058] The two-dimensional cognitive state space is divided into four cognitive state regions: the first cognitive region corresponds to a physical consistency index below the first preset threshold and a cognitive uncertainty entropy above the second preset threshold, representing a state of dual physical and cognitive imbalance in the model. This region is the most critical cognitive state requiring intervention. A physical consistency index below the first preset threshold indicates a significant discrepancy between the 3D defect structure predicted by the 3D convolutional neural network model and the results derived from the fracture mechanics theory model of carbon fiber composite materials, meaning the model's judgment of the defect does not conform to the physical laws of material failure. A cognitive uncertainty entropy above the second preset threshold indicates a large dispersion in the model's predictions of the same defect at different imaging angles, meaning the model lacks confidence in its own predictions. Both indicators deviating from the normal range simultaneously means the model neither understands the physical laws (low physical consistency) nor has confidence in its own predictions (high cognitive uncertainty), resulting in a state of comprehensive cognitive confusion. For example, for a layered defect, the AI ​​model predicts it is located at a depth of 3 mm, while the physical model deduces that the excitation response feature should originate from a pore at a depth of 1 mm; simultaneously, the prediction results from four input angles contradict each other: 0 degrees predicts layering, 90 degrees predicts no defect, and 180 degrees predicts a crack. At this point, the model is in the first cognitive region and requires the strongest intervention.

[0059] The second cognitive region corresponds to a physical consistency index below the first preset threshold and a cognitive uncertainty entropy not exceeding the second preset threshold, indicating that the model is in a state of insufficient understanding of physical laws. This region is characterized by a low degree of adherence to physical laws, but a relatively high degree of confidence in its own predictions. A physical consistency index below the first preset threshold indicates a discrepancy between the AI ​​model's predictions and the fracture mechanics theory model; the model's output of defect morphology, location, or size does not conform to physical laws. A cognitive uncertainty entropy not exceeding the second preset threshold indicates good consistency among predictions from different angles; the model is actually quite confident in its predictions that do not conform to physical laws. For example, for a certain impact damage, the AI ​​model consistently predicts a circular layer with a diameter of 10 mm from four angles, with confidence levels all above 0.9 and a very low cognitive uncertainty entropy. However, the fracture mechanics theory model, based on surface thermal response characteristics, infers that the damage should be an elliptical layer with its major axis along the fiber direction, a significant difference from the AI ​​prediction. In this case, the model is in the second cognitive region, and its problem is "persistently making mistakes," requiring stronger constraints from physical laws.

[0060] The third cognitive region corresponds to a physical consistency index no lower than the first preset threshold and a cognitive uncertainty entropy higher than the second preset threshold, indicating that the model is in a state of insufficient understanding of data distribution. The characteristic of this region is that the model follows physical laws relatively well, but has a low degree of confidence in its own predictions. A physical consistency index no lower than the first preset threshold indicates that the AI ​​model's predictions are basically consistent with the fracture mechanics theoretical model and conform to physical laws. A cognitive uncertainty entropy higher than the second preset threshold indicates that there are significant discrepancies between the prediction results from different angles, and the model's judgment of the same defect is inconsistent from different perspectives. For example, for a certain micropore defect, the physical model's deduction and the AI ​​prediction are basically consistent in terms of defect location and size, with a high physical consistency index; however, the probability of the defect existing at 0 degrees is predicted to be 0.9, at 90 degrees only 0.3, and at 180 degrees 0.6, showing large dispersion between angles, indicating that the model is not confident in its own predictions. At this point, the model is in the third cognitive region, and its problem is "knowing the rules but not seeing them clearly," requiring more real detection data to improve its recognition ability.

[0061] The fourth cognitive region corresponds to a physical consistency index no lower than the first preset threshold and a cognitive uncertainty entropy no higher than the second preset threshold, indicating that the model is in a state of cognitive equilibrium. This region represents the ideal state of model cognition. A physical consistency index no lower than the first preset threshold indicates that the AI ​​model's predictions are highly consistent with the fracture mechanics theory model and conform to physical laws. A cognitive uncertainty entropy no higher than the second preset threshold indicates a high degree of consistency between predictions from different angles, demonstrating the model's confidence in its conforming predictions. For example, for a typical layered defect, the AI ​​model stably predicts from all four angles a circular layer with a diameter of 8 mm and a depth of 2 mm, with confidence levels all above 0.95. Simultaneously, this prediction perfectly matches the forward derivation result based on fracture mechanics theory, with a positive and relatively large phase interference term. At this point, the model is in the fourth cognitive region, in a state of cognitive equilibrium, requiring no additional intervention; maintaining the current learning parameters is sufficient.

[0062] The first and second preset thresholds are set based on the following: The first preset threshold is used to determine whether the physical consistency index meets the standard. Its setting references the industrial acceptance standards in the field of non-destructive testing of carbon fiber composite materials. According to the testing specifications for aerospace carbon fiber composite materials, the allowable deviation for the depth of delamination defects is no more than 0.5 mm, the area coverage is no more than 2%, and the pore volume fraction is no more than 1.0%. Accordingly, the first preset threshold can be set to make the physical consistency index correspond to the upper limit of the above allowable deviation range. That is, when the deviation between the AI ​​prediction and the physical model exceeds the allowable range of the industrial standard, it is determined that the physical consistency index is lower than the first preset threshold. The second preset threshold is used to determine whether the cognitive uncertainty entropy is too high. Its setting adopts a statistical method, based on the distribution of cognitive uncertainty entropy in historical testing data, taking the percentile as the threshold. For example, the position where the entropy value exceeds 85% of historical samples is set as the second preset threshold. Alternatively, cross-validation can be used to test the detection accuracy corresponding to different thresholds on the validation set, selecting the entropy value that causes the accuracy to begin to decrease significantly as the second preset threshold. The specific values ​​of the two thresholds are not fixed, but are dynamically calibrated as the system runs and more data accumulates. However, once set, they remain stable in the current testing batch and are used for the division of state regions.

[0063] The cognitive state regulation module is also used to: when the state point is located in the first cognitive region, perform collaborative regulation operations to simultaneously enhance the weight of the physical constraint regularization term during the training process of the 3D convolutional neural network model and increase the proportion of newly introduced real detection data in incremental learning. This operation is for models in a state of dual physical and cognitive imbalance. At this time, the model deviates from physical laws and lacks confidence in its own predictions, requiring strong intervention from both physical and data dimensions. The enhancement of the weight of the physical constraint regularization term is achieved by adjusting the coefficient of the physical consistency term in the loss function. For example, the original loss function is the data fitting loss plus the physical constraint regularization term (coefficient λ). During collaborative regulation, λ is increased from 0.1 to 0.5, making the model training pay more attention to the constraints that conform to fracture mechanics theory, forcing the prediction results to move closer to physical laws. At the same time, the proportion of newly introduced real detection data in incremental learning is increased from the default 5% to 15%, that is, the proportion of real defect samples in each batch of training data increases, allowing the model to have more contact with the defect morphology in actual production and reducing the dependence on simulation data. For example, originally every 100 training samples contained 5 real defect samples and 95 simulated samples, but after adjustment, this was changed to 15 real samples and 85 simulated samples. Simultaneous augmentation allows the model to strengthen physical constraints while correcting biases with more real data, quickly escaping a state of cognitive confusion.

[0064] When the state point is located in the second cognitive region, a physical orientation adjustment operation is performed, only increasing the weight of the physical constraint regularization term while maintaining the current ratio of real detection data. This operation is for models that are in a state of insufficient understanding of physical laws. At this time, the model's prediction consistency across different angles is good (low cognitive uncertainty), but the prediction results do not conform to physical laws, indicating that the model is "consistently making mistakes." The root cause of the problem lies in the lack of physical constraints rather than insufficient data. Therefore, only the weight of the physical constraint regularization term is increased without changing the data ratio. In specific implementation, the coefficient of the physical constraint regularization term in the loss function is increased from the current value (e.g., 0.1) to 0.3 or 0.4. The specific magnitude is dynamically determined based on the degree to which the physical consistency index deviates from the threshold; the greater the deviation, the greater the increase. For example, if the physical consistency index is 20% lower than the first preset threshold, the regularization term coefficient is increased by 50%. The ratio of real detection data remains unchanged (e.g., 5%) because the model already has sufficient grasp of the data and does not require additional data. This operation enables the model to more strictly follow the laws of fracture mechanics during training, gradually correcting its prediction habits that do not conform to physics.

[0065] When the state point is located in the third cognitive region, a data-oriented adjustment operation is performed, increasing only the proportion of newly introduced real detection data in incremental learning while maintaining the current weight of the physical constraint regularization term. This operation is for models in a state of insufficient data distribution cognition. At this time, the model's predictions conform to physical laws (high physical consistency), but the prediction results from different perspectives differ greatly (high cognitive uncertainty), indicating that the model "knows the rules but cannot see them clearly." The root cause of the problem lies in the lack of sufficient real defect samples, leading to insufficient identification ability. Therefore, only the proportion of real detection data is increased without changing the strength of physical constraints. In specific implementation, the proportion of real samples in incremental learning is increased from the current value (e.g., 5%) to 10% or 12%, with the specific magnitude determined by the degree to which the cognitive uncertainty entropy exceeds the second preset threshold; the higher the entropy value, the greater the increase. For example, if the cognitive uncertainty entropy exceeds the threshold of 30%, the proportion is increased to 8%. The weight of the physical constraint regularization term remains unchanged (e.g., still 0.1). This operation allows the model to learn more accurate discrimination boundaries through more real defect samples, improving the consistency of multi-view predictions without interfering with its existing cognitive foundation that conforms to physical laws.

[0066] When the state point is located in the fourth cognitive region, a steady-state maintenance operation is performed, maintaining the current weights of the physical constraint regularization term and the current ratio of real detection data unchanged. This operation is for models in a cognitive equilibrium state. At this point, the model's predictions conform to physical laws and are consistent across multiple perspectives, representing an ideal working state. No adjustments are needed; the existing training parameters remain unchanged, allowing the model to continue running stably. For example, the weights of the physical constraint regularization term are maintained at 0.1, and the ratio of real detection data is maintained at 5%. The system continues incremental learning according to the original strategy until the next cognitive state evaluation detects a deviation. This operation ensures that the model steadily accumulates knowledge in a cognitive equilibrium state, avoiding unnecessary perturbations.

[0067] Through the above complete design, the present invention achieves differentiated handling of different energy consumption bottlenecks: problems of equipment itself and pipeline pressure loss are addressed through alarm prompts for manual intervention; problems of multi-factor coupling are addressed through comprehensive diagnosis and auxiliary analysis; and problems of improper operation strategies are addressed through closed-loop control and automatic optimization, thus forming a complete closed loop from monitoring, diagnosis to handling.

[0068] The above-mentioned models or function formulas are all dimensionless and numerical calculations. The models or function formulas are obtained by software simulation based on a large amount of collected data to obtain the most recent real situation. The preset parameters in the models or function formulas are set by those skilled in the art according to the actual situation.

[0069] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0070] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0071] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0072] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A defect location system for carbon fiber bicycle frame production, characterized in that, include: The image acquisition module is used to acquire two-dimensional detection images, which are the surface physical field response images of the carbon fiber bicycle frame under test under a preset excitation load. The 3D inversion module is used to process the 2D detection image based on the pre-trained 3D convolutional neural network model and output the 3D structural information of the internal defects of the carbon fiber bicycle frame to be tested. The 3D convolutional neural network model is incrementally trained from source domain simulation data to target domain real detection data through transfer learning. The incremental transfer module is used to execute the incremental learning process. During the incremental learning process, the ratio between the newly introduced real detection data and the historical simulation data used for pre-training is adjusted in real time according to the corresponding differentiated incremental learning control strategy generated by the cognitive state control module. The cognitive metrics module is used to calculate the physical consistency index and cognitive uncertainty entropy of a 3D convolutional neural network model when outputting 3D structural information in real time. The cognitive state regulation module is used to construct a two-dimensional cognitive state space composed of physical consistency index and cognitive uncertainty entropy. It maps the current cognitive state of the three-dimensional convolutional neural network model to a state point in the two-dimensional cognitive state space and generates corresponding differentiated incremental learning regulation strategies based on the cognitive state region where the state point is located.

2. The defect location system for carbon fiber bicycle frame production according to claim 1, characterized in that, The image acquisition module includes: The multi-segment heterogeneous frequency thermal excitation unit is used to apply thermal pulse excitation of different frequencies and durations according to the curvature radius and wall thickness distribution of different parts of the carbon fiber bicycle frame under test. The multi-angle synchronous imaging unit includes at least three infrared thermal imagers arranged circumferentially around the carbon fiber bicycle frame under test. The imaging optical axes of the at least three infrared thermal imagers intersect at the central axis of the carbon fiber bicycle frame under test, and are used to synchronously acquire image sequences of the surface temperature field evolution of the carbon fiber bicycle frame under test at different angles during thermal pulse excitation. The layup orientation compensation unit is used to acquire the carbon fiber layup orientation information of the carbon fiber bicycle frame under test in the image acquisition area, and to perform anisotropic thermal diffusion compensation on the surface temperature field evolution image sequence based on the layup orientation information, generating a standardized thermal response image after eliminating the influence of layup orientation, which is then used as a two-dimensional detection image input to the three-dimensional inversion module.

3. A defect location system for carbon fiber bicycle frame production according to claim 2, characterized in that, The multi-segment heterogeneous frequency thermal excitation unit is specifically used for: The radius of curvature ρ and wall thickness d of different parts of the carbon fiber bicycle frame under test are obtained, and the thermal pulse excitation parameters are determined based on the radius of curvature ρ and wall thickness d. The thermal pulse excitation parameters include the pulse frequency f and the pulse duration t. The pulse frequency f satisfies f=k1 / (ρ*d)+f0 with respect to the radius of curvature ρ and the wall thickness d, and the pulse duration t satisfies t=k2*ρ*d+t0 with respect to the radius of curvature ρ and the wall thickness d. Where k1 and k2 are proportionality coefficients related to the thermal diffusivity of carbon fiber composites, and f0 and t0 are inherent parameters of the excitation system.

4. A defect location system for carbon fiber bicycle frame production according to claim 3, characterized in that, The ply orientation compensation unit is specifically used for: The carbon fiber layup orientation angle θ of the carbon fiber bicycle frame under test in the image acquisition area is obtained. The layup orientation angle θ is defined as the angle between the carbon fiber axis and the preset reference direction. A thermal diffusion anisotropy model of carbon fiber composites was established to describe the propagation law of heat in carbon fiber composites. The thermal diffusion coefficient along the carbon fiber axis is 5 to 10 times that perpendicular to the carbon fiber axis. The isothermal surface of heat propagation is ellipsoidal, and the major axis of the ellipsoid is consistent with the carbon fiber axis. In the plane perpendicular to the fiber, the diffusion rate of heat along the warp and weft directions is faster than the diffusion rate along the 45-degree direction. Based on the anisotropic thermal diffusion model and the layup orientation angle θ, a two-dimensional point diffusion function h(x,y) is constructed. The point diffusion function h(x,y) is: ； in, , σx is positively correlated with the thermal diffusivity along the fiber axis, and σy is positively correlated with the thermal diffusivity perpendicular to the fiber axis, with the value of σx / σy ranging from 5 to 10. The image of each frame in the surface temperature field evolution image sequence is deconvolved with the point spread function h(x,y) to eliminate the spatial broadening effect of the temperature field caused by the anisotropic thermal diffusion of carbon fibers and restore the true geometric shape of the defect. The standardized thermal response image after deconvolution is output and used as the input of the two-dimensional detection image to the three-dimensional inversion module.

5. A defect location system for carbon fiber bicycle frame production according to claim 4, characterized in that, 3D convolutional neural network models are used for: Receive two-dimensional detection images from multiple angles acquired by the image acquisition module, including at least three different viewpoints distributed circumferentially around the carbon fiber bicycle frame under test; Feature extraction is performed on the two-dimensional detection image from each viewpoint to generate a two-dimensional feature map for the corresponding viewpoint. Feature extraction is achieved through convolutional layers with shared weights. Two-dimensional feature maps from multiple perspectives are mapped to three-dimensional space through projection transformation to generate an initial three-dimensional feature volume. The projection transformation is based on the response characteristics of carbon fiber composite material to a preset excitation load. The response characteristics include the correlation between the X-ray attenuation coefficient or thermal diffusivity coefficient and the material density. The X-ray attenuation coefficient is suitable for X-ray excitation load scenarios, and the thermal diffusivity coefficient is suitable for thermal pulse excitation load scenarios. The correlation is pre-calibrated by the physical properties of carbon fiber composite material. Perform 3D convolution operation on the initial 3D feature volume, fuse multi-view information and eliminate reconstruction ambiguity in the projection process to generate an enhanced 3D feature volume; The enhanced 3D feature volume is upsampled layer by layer to a preset 3D spatial resolution, and a 3D voxel representation is generated through the output layer. The value of each voxel in the 3D voxel representation represents the probability value of the existence of internal defects at that spatial location. The output is a three-dimensional voxel representation of the three-dimensional structural information of the internal defects of the carbon fiber bicycle frame under test.

6. A defect location system for carbon fiber bicycle frame production according to claim 5, characterized in that, The physical consistency index is calculated as follows: Obtain the 3D structural information of the defect predicted by the 3D convolutional neural network model, and use it as the first prediction vector; The three-dimensional structural information of the defect obtained by forward deduction based on the fracture mechanics theoretical model of carbon fiber composite material is used as the second prediction vector. The fracture mechanics theoretical model is established based on the excitation response characteristics and material property parameters in the two-dimensional detection image. The first and second prediction vectors are mapped to the complex space and expressed in complex form, where the modulus of the complex number represents the defect amplitude information and the argument of the complex number represents the defect phase information. Calculate the complex interference term between the first prediction vector and the second prediction vector. The complex interference term is defined as the inner product of the two vectors in the complex space minus half the sum of the squares of their moduli. The physical consistency index is determined based on the sign and magnitude of the complex interfering terms. When the complex interfering terms are positive, the physical consistency index is positively correlated with the magnitude of the complex interfering terms, indicating that the predictions of the physical model and the AI ​​model are in phase and reinforce each other. When the complex interfering terms are negative, the physical consistency index is negatively correlated with the absolute value of the complex interfering terms, indicating that the predictions of the physical model and the AI ​​model are in opposite phase and weaken each other. When the complex interfering terms are zero, the physical consistency index is zero, indicating that the two are orthogonal and have no interference.

7. A defect location system for carbon fiber bicycle frame production according to claim 6, characterized in that, The method for calculating the entropy of cognitive uncertainty is as follows: Obtain the prediction results of the three-dimensional convolutional neural network model for the same defect under multiple different imaging angles. The prediction results under each imaging angle include the three-dimensional structural information of the defect and its confidence distribution. The prediction results at each imaging angle are subjected to probability transformation to convert them into a probability distribution function at that angle. The probability distribution function represents the probability value of the existence of defects at each spatial location. Perform a Fourier transform on each probability distribution function to obtain its frequency domain representation, and extract the phase angle at the point of maximum amplitude in the frequency domain representation as the phase feature of the probability distribution function at that angle; Calculate the normalized cross-correlation similarity between the probability distribution functions corresponding to any two different imaging angles. The normalized cross-correlation similarity is obtained by integrating the product of the two probability distribution functions and then dividing by the product of the square roots of their respective square integrals. Complex interferometric coefficients are constructed based on the absolute value of the normalized cross-correlation similarity and the phase difference between the two imaging angles. The modulus of the complex interferometric coefficient is the absolute value of the normalized cross-correlation similarity, and the argument of the complex interferometric coefficient is the phase difference. Construct a complex interference matrix composed of complex interference coefficients. The complex interference matrix is ​​a square matrix in which the number of rows and columns are equal to the number of imaging angles. By performing eigenvalue decomposition on the complex interference matrix, multiple complex eigenvalues ​​are obtained; The complex entropy value of the complex interference matrix is ​​calculated. The real part of the complex entropy value is determined by the weighted sum of the negative logarithms after modulus normalization of each complex eigenvalue, and the imaginary part of the complex entropy value is determined by the average argument of each complex eigenvalue. The modulus of the complex entropy value is used as the cognitive uncertainty entropy.

8. A defect location system for carbon fiber bicycle frame production according to claim 7, characterized in that, The two-dimensional cognitive state space is divided into four cognitive state regions: The first cognitive region corresponds to a physical consistency index that is lower than the first preset threshold and a cognitive uncertainty entropy that is higher than the second preset threshold, indicating that the model is in a state of dual physical and cognitive imbalance. The second cognitive region corresponds to a physical consistency index that is lower than the first preset threshold and a cognitive uncertainty entropy that is not higher than the second preset threshold, indicating that the model is in a state of insufficient physical law cognition. The third cognitive region corresponds to a physical consistency index that is not lower than the first preset threshold and a cognitive uncertainty entropy that is higher than the second preset threshold, indicating that the model is in a state of insufficient data distribution cognition. The fourth cognitive region corresponds to a physical consistency index that is not lower than the first preset threshold and a cognitive uncertainty entropy that is not higher than the second preset threshold, indicating that the model is in a state of cognitive equilibrium.

9. A defect location system for carbon fiber bicycle frame production according to claim 8, characterized in that, The cognitive state regulation module is also used for: When the state point is located in the first cognitive region, a collaborative regulation operation is performed to simultaneously enhance the weight of the physical constraint regularization term in the training process of the 3D convolutional neural network model and increase the proportion of newly introduced real detection data in incremental learning. When the state point is located in the second cognitive region, a physical orientation control operation is performed, which only increases the weight of the physical constraint regularization term and maintains the current real detection data ratio. When the state point is located in the third cognitive region, data-oriented adjustment operations are performed, which only increase the proportion of real detection data and maintain the weight of the current physical constraint regularization term. When the state point is located in the fourth cognitive region, a steady-state maintenance operation is performed to maintain the current weight of the physical constraint regularization term and the current ratio of the actual detection data unchanged.

Images