An automatic recognition method for HTCC ceramic defects based on industrial model assistance

Abstract

The present invention relates to the field of intelligent manufacturing technology, and more particularly to an industrial model-assisted automatic defect identification method for HTCC ceramics. The method comprises the following steps: 1: real-time acquisition of process parameters during the HTCC production process, including sintering temperature gradient, lamination pressure, and green body moisture content; 2: non-contact dielectric property scanning of the sintered HTCC substrate on a production line conveyor belt; 3: based on the coordinates of the dielectric anomaly region, controlling a circularly polarized light source to illuminate the target region at an incident angle that suppresses reflection from the metallization layer; 4: inputting the dielectric anomaly frequency band feature vector and the defect image texture features from step 3 into a cross-modal attention network to generate a fused feature map; and 5: establishing a correlation mapping library between defect type and process deviation type. Through precise real-time monitoring, intelligent defect identification and classification, automated process adjustment, and a continuous optimization feedback mechanism, the method significantly improves the efficiency, quality, and intelligence level of the ceramic production process.

Description

Technical Field

[0001] The present invention relates to the field of intelligent manufacturing technology, and in particular to an automatic identification method for HTCC ceramic defects based on industrial model assistance. Background Art

[0002] High-temperature co-fired ceramics (HTCC) are widely used in high-end fields such as aerospace high-frequency devices and high-power electronic packaging due to their excellent high-temperature resistance, high mechanical strength, and multi-layer wiring capabilities. However, the HTCC production process is complex, involving key steps such as tape casting, lamination, and co-firing. Fluctuations in process parameters (such as abnormal sintering temperature gradients and uneven lamination pressure) can easily lead to defects such as internal microcracks, interlayer voids, and metallization layer delamination. These defects can cause serious problems such as signal distortion and thermal failure when the device operates at high frequencies, posing a significant threat to the reliability of electronic systems on board spacecraft. Current technologies for HTCC ceramic defect detection have the following limitations:

[0003] Existing technologies mostly focus on offline inspection of finished products (such as X-ray tomography and ultrasonic testing). While these methods can identify defects that have already occurred, they cannot trace the relationship between defects and process parameters during production (such as temperature fluctuations in the sintering zone and excessive moisture content in the green body). For example, a certain type of HTCC substrate for aerospace applications developed batches of interlayer cracks due to uncontrolled heating rates in the temperature zones. Traditional methods could only eliminate scrap products but could not pinpoint the root cause of the process, leading to the recurrence of similar defects.

[0004] During the HTCC sintering process, differences in the thermal expansion coefficients of interlayer materials trigger dynamic stress accumulation, ultimately forming microcracks. Existing X-ray inspection requires production interruption and cannot capture defect evolution in real time. Surface optical imaging methods are limited by the high reflectivity of the metallization layer, making it difficult to reliably capture internal defect signatures. The failure of a satellite communication module during service due to the propagation of internal microcracks demonstrates that the lack of online dynamic monitoring directly threatens the long-life reliability of high-value components. Summary of the Invention

[0005] Based on the above objectives, the present invention provides an automatic identification method for HTCC ceramic defects based on industrial model assistance, comprising the following steps:

[0006] Step 1: Real-time collection of HTCC production process parameters including sintering temperature gradient, lamination pressure, and green body moisture content. A spatiotemporal baseline model of the process parameters is constructed using a long-short-term memory network. The deviation between the current parameters and the baseline model is dynamically calculated. When the deviation exceeds a dynamic threshold adaptively set based on historical stability data of the production line, a process deviation type label is generated.

[0007] Step 2: Perform a non-contact dielectric property scan on the sintered HTCC substrate on the production line conveyor. Dynamically select the dielectric scanning frequency band based on the process deviation type label in step 1. Obtain the dielectric constant distribution map using the resonant cavity perturbation method. Calculate the local variance and mark the coordinates of the dielectric anomaly area. Map these coordinates to the optical detection coordinate system.

[0008] Step 3: Based on the coordinates of the dielectric anomaly area, control the circularly polarized light source to illuminate the target area at an incident angle that suppresses reflection from the metallization layer. Dynamically match the near-infrared imaging wavelength based on the dielectric scanning frequency band, and subtract the light intensity of the normal area through the background difference method to generate a high-contrast image of the local defect.

[0009] Step 4: Input the dielectric anomaly frequency band feature vector and the defect image texture features from step 3 into the cross-modal attention network to generate a fused feature map. Activate the corresponding defect classification branch based on the process deviation type label, and use a generative adversarial network to separate the defect features from the process background texture, outputting the defect type and severity level.

[0010] Step 5: Establish an association mapping library between defect types and process deviation types. When the same defect type appears a preset number of times consecutively, automatically adjust the sintering furnace temperature control parameters, verify the change in defect rate after parameter adjustment through the digital twin model, update the dynamic threshold of process deviation, and form a closed-loop feedback loop.

[0011] Preferably, the method for determining the adaptive dynamic threshold based on historical stability data of the production line in step 1 includes:

[0012] Collect temperature gradient data, lamination pressure time series data, and green body moisture content detection data in each temperature zone during the continuous production cycle of the production line to build a historical database of process parameters;

[0013] Divide historical data by production batches, calculate the standard deviation and mean shift of parameters within each batch, and generate stability evaluation indicators;

[0014] Based on the stability evaluation index corresponding to the current production batch, a sliding window algorithm is used to dynamically adjust the deviation judgment threshold: when the stability evaluation index is lower than the preset critical value, the tolerance range of the deviation judgment threshold is expanded; when the index is higher than the critical value, the threshold range is narrowed;

[0015] The generation of the process deviation type label includes: matching the current parameter deviation pattern with a predefined process abnormality pattern library, wherein the pattern library is established by clustering historical abnormality data, and each pattern contains a deviation weight combination of temperature, pressure, and moisture content.

[0016] Preferably, the method for dynamically selecting the dielectric scanning frequency band according to the process deviation type label in step 2 includes:

[0017] Construct a dielectric frequency band-defect type sensitivity mapping table: experimentally measure the dielectric parameter change rate of various defects at different frequency bands and calculate the frequency band's response intensity to the defect;

[0018] Based on the process deviation type label in step 1, select the top N frequency bands with the highest response intensity from the mapping table as the scanning frequency bands. N is dynamically adjusted based on the real-time requirements of the production line inspection.

[0019] The mapping method of the dielectric abnormal area coordinates includes: setting a reference positioning mark on the conveyor belt, obtaining the real-time position of the substrate through a laser rangefinder, establishing an affine transformation relationship between the dielectric scanning coordinate system and the optical imaging coordinate system, and achieving submillimeter spatial alignment.

[0020] Preferably, the method for adjusting the incident angle of the circularly polarized light source to suppress reflection of the metallized layer in step 3 includes:

[0021] Collect reflectivity data of the metallization layer on the surface of the HTCC substrate, train the light source angle optimization model through historical imaging samples, and output the optimal incident angle range for suppressing specular reflection;

[0022] The incident angle is corrected in real time based on the surface roughness detection data of the current substrate. The angle between the incident angle and the normal is reduced when the roughness is higher.

[0023] The implementation of the background difference method includes: collecting the light intensity distribution template of the normal area, updating the template through a dynamic weighted average algorithm, and deducting the template data from the target image to retain the difference signal of the defect area.

[0024] Preferably, the method for constructing the cross-modal attention network by inputting the dielectric anomaly frequency band feature vector and the defect image texture feature into the cross-modal attention network in step 4 includes:

[0025] Normalize the dielectric frequency band feature vector according to the frequency band sensitivity weight and use it as the key value of the attention mechanism;

[0026] Extract the local binary pattern texture features of the defect image and use them as query values after dimension reduction through the convolution layer;

[0027] A multi-head attention mechanism is used to calculate the cross-modal weight matrix and suppress feature channels that are irrelevant to the current process deviation type.

[0028] The constraints of the adversarial generative network include: limiting the defect morphology distribution output by the generator according to the process deviation type label, and the discriminator receiving real defect samples and synthetic samples generated by process parameter perturbations for adversarial training.

[0029] Preferably, the training data generation method for separating defect features and process background texture using a generative adversarial network in step 4 includes:

[0030] When a new type of defect is detected, a simulated defect texture is superimposed on the normal sample according to its associated process deviation type label;

[0031] The method for generating the simulated defect texture includes: simulating the stress distribution caused by the deviation of the process parameters through finite element analysis to generate the morphological characteristics of the crack or hole;

[0032] The simulated defect images are spatially aligned with the real dielectric scanning data to construct an augmented training dataset.

[0033] Preferably, the method of verifying the change in defect rate after parameter adjustment by the digital twin model in step 5 includes:

[0034] Build a thermal-mechanical coupling simulation model for the HTCC sintering process, input the current process parameters and the parameter combination to be adjusted;

[0035] Simulate the expansion coefficient difference and thermal stress distribution between layers during sintering to predict the probability and distribution of defects;

[0036] When the reduction in the simulated defect rate does not meet expectations, iteratively adjust the parameters and re-simulate until the verification conditions are met;

[0037] The method for determining the preset number of times includes: setting different thresholds according to the severity level of the defect, and the higher the severity, the fewer consecutive occurrences are required to trigger parameter adjustment.

[0038] Preferably, the method for establishing the dielectric frequency band-defect type sensitivity mapping table includes:

[0039] In samples with known defects, the correlation coefficient between the dielectric constant change rate and the defect size in each frequency band is measured;

[0040] Use principal component analysis to screen the frequency band combination that contributes most to defect classification;

[0041] The comprehensive sensitivity score of the frequency band is calculated based on the correlation coefficient and the principal component weight. The higher the score, the higher the scanning priority.

[0042] Preferably, the method for acquiring the surface roughness detection data includes:

[0043] Before dielectric scanning, the substrate surface roughness profile was measured by laser triangulation;

[0044] Calculate the surface roughness grade based on the peak-to-valley height difference and root mean square value of the profile data;

[0045] The incident angle correction rule is: for each increase in the roughness level, the angle between the incident angle and the normal line decreases by a preset angle step.

[0046] Preferably, the calibration method of the thermal-mechanical coupling simulation model includes:

[0047] Regularly collect temperature gradient data and interlaminar stress detection data during the actual sintering process and compare them with the simulation results;

[0048] When the error exceeds the allowable range, the material parameter boundary conditions of the simulation model are updated through the back propagation algorithm;

[0049] The material parameters include thermal conductivity, thermal expansion coefficient and viscoelastic modulus of the HTCC green body.

[0050] Beneficial effects of the present invention:

[0051] By collecting and analyzing process parameters in real time, anomalies in the production process can be promptly detected and corrected, reducing the production of defective products. Furthermore, the dynamic adjustment mechanism for process parameters ensures that the quality of ceramic products is consistently maintained at optimal levels. Leveraging dielectric property scanning and optical imaging technology, this method accurately locates and identifies defects on HTCC ceramic substrates and effectively classifies these defects using a cross-modal attention network. This provides sufficient data support for subsequent defect processing and analysis. By analyzing the correlation between process parameters and defect types, this method automatically adjusts relevant production parameters when defects occur repeatedly, reducing defects caused by process issues. This continuous optimization process ensures production line stability and long-term product quality reliability. Furthermore, automated data collection, analysis, and process adjustments reduce the need for manual intervention and improve the automation level of the production line. The combination of a digital twin model and a closed-loop feedback mechanism makes the entire production process more intelligent and adaptive. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only for the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0053] Figure 1 is a flow chart of the steps of the method of the present invention;

[0054] Figure 2 This is a flowchart of the method for verifying the change in defect rate after parameter adjustment through a digital twin model in step 5 of the method of the present invention;

[0055] Figure 3 The figure is a flowchart of the steps of the method for establishing the dielectric frequency band-defect type sensitivity mapping table in the method of the present invention. DETAILED DESCRIPTION

[0056] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. It is also noted that, to provide a more detailed description, the following embodiments are best and preferred embodiments, and those skilled in the art may employ alternative methods for implementing certain known technologies. Furthermore, the accompanying drawings are intended only to provide a more detailed description of the embodiments and are not intended to limit the present invention.

[0057] See Figure 1-Figure 3 An embodiment of the present invention provides an automatic defect identification method for HTCC ceramics, assisted by industrial models. During the production process, process parameters such as sintering temperature gradient, lamination pressure, and green body moisture content are collected in real time. This data is analyzed using a spatiotemporal baseline model constructed using a long short-term memory (LSTM) network. The LSTM model processes time series data and dynamically calculates the deviation between the current parameters and the baseline model. If the deviation exceeds a dynamic threshold adaptively based on historical production line stability data, the system generates a process deviation type label. This process enables early detection of production anomalies and provides a basis for subsequent adjustments.

[0058] Based on the process deviation type label generated in step 1, select the appropriate scanning frequency band for dielectric properties scanning. Use the resonant cavity perturbation method to obtain the dielectric constant distribution map of the HTCC substrate, and calculate the local variance to identify areas of dielectric anomaly. By mapping the coordinates of these anomaly areas to the optical inspection coordinate system, precise spatial positioning is provided for subsequent defect detection and image generation.

[0059] After scanning an abnormal area, the incident angle of the circularly polarized light source is controlled to suppress interference caused by reflections from the metallization layer. Using background subtraction, the light intensity of the normal area is subtracted to generate a high-contrast image of the local defect. This process makes the defect area more prominent, facilitating subsequent image processing and defect identification.

[0060] The defect image texture features generated in step 3 are fed into a cross-modal attention network along with the feature vectors of the dielectric anomaly frequency band. By calculating the weighted weights of the cross-modal features, this network effectively extracts important information relevant to defect classification. A generative adversarial network (GAN) further separates the defect features from the process background texture, ultimately outputting the defect type and severity level.

[0061] Based on a mapping library of defect types and process deviation types, the system automatically adjusts the sintering furnace's temperature control parameters when the same defect type occurs a preset number of times consecutively. The results of these adjustments are verified using a digital twin model, simulating how the new process parameters affect the defect rate. The system also dynamically updates the process deviation threshold based on changes in the defect rate, creating a closed-loop feedback loop. This mechanism enables continuous process optimization throughout the production process, ensuring consistent product quality.

[0062] In one possible implementation, the system first collects temperature gradient data for each temperature zone, lamination pressure time series data, and green body moisture content data throughout the continuous production cycle of the production line. This data is used to build a historical database of process parameters. This historical data is then used to conduct retrospective analysis and modeling of the production process, ensuring that potential process fluctuations and abnormal trends are captured.

[0063] Next, the system divides historical data into production batches and calculates the standard deviation and mean shift of each process parameter within each batch. Using this data, it generates stability evaluation indicators that quantify the stability of each batch. The standard deviation reflects the volatility of the parameters, while the mean shift shows the long-term trend of each process parameter. As an important parameter for measuring the current production status, the stability evaluation indicator directly influences the strategy for dynamically adjusting the deviation threshold.

[0064] Based on the generated stability evaluation index, a sliding window algorithm is used to dynamically adjust the deviation threshold. Specifically, when the stability evaluation index falls below the preset critical value, the system automatically expands the tolerance range of the deviation threshold, allowing for greater parameter deviation during production. When the stability evaluation index rises above the critical value, the system narrows the threshold range, requiring process parameters to remain more strictly within the standard range. This dynamic adjustment mechanism can flexibly adapt to the production line status based on the different characteristics of production batches, thereby improving the system's stability and adaptability.

[0065] After dynamically adjusting the deviation threshold, the system matches the deviation pattern of the current process parameters with a predefined library of process anomaly patterns based on the deviation of the process parameters during the current production process. This library is built by clustering historical anomaly data, with each pattern containing a weighted combination of deviations for process parameters such as temperature, pressure, and moisture content. When the current process parameters match a specific anomaly pattern, the system generates a corresponding process deviation type label, which serves as an important basis for subsequent defect detection and classification.

[0066] In one possible implementation, the rate of change of dielectric parameters for various types of defects at different frequency bands is first measured experimentally. Each defect exhibits different response strengths in different frequency bands, with some frequency bands being more sensitive to certain defects and others to others. Using this experimental data, a sensitivity mapping table can be established between dielectric frequency bands and defect types. This table records the response strength of each frequency band to various defects, providing a basis for subsequent selection of scanning frequency bands.

[0067] Based on the process deviation type label, the system can select the frequency band that is most sensitive to the current process deviation pattern from the mapping table. The process deviation type label is generated by the previous step and represents the abnormal characteristics of the process in the current production process. By looking up the response strength of the corresponding frequency band in the mapping table, the system will select the top N frequency bands with the highest response strength as the scanning frequency bands. The value of N is dynamically adjusted according to the real-time requirements of production line detection: if a faster response to defects is required, fewer frequency bands (for example, the first two frequency bands) may be selected, and if more precise identification is required, more frequency bands may be selected to obtain more comprehensive detection data.

[0068] To ensure that the dielectric scanning results accurately match the defect locations in the actual production process, the system requires precise spatial alignment. To achieve this, reference positioning marks are placed on the conveyor belt to track the position of the substrate in real time. A laser rangefinder is used to obtain the real-time position of the substrate, and the positional relationship between the reference marks and the substrate is used to establish an affine transformation relationship between the dielectric scanning coordinate system and the optical imaging coordinate system. This step ensures the accurate connection between the scan data and the optical imaging data, thereby achieving submillimeter spatial alignment, allowing the inspection results to accurately indicate the specific location of the defect.

[0069] In one possible implementation, reflectance data is first collected from the metallization layer on the HTCC substrate surface at different incident angles. The reflectance characteristics of the metallization layer vary with the incident angle. In particular, at certain angles, reflected light may cause specular reflections in the image, affecting the detection of defective areas. Therefore, by collecting reflectance data at different angles, the system can construct a light source angle optimization model. Based on historical imaging samples, this model is trained to predict the optimal incident angle range for suppressing specular reflections, thereby optimizing the incident angle to reduce specular reflection interference from the metallization layer.

[0070] The surface roughness of HTCC substrates directly affects the reflective properties of light. When the substrate surface roughness is high, surface irregularities lead to uneven light reflection, necessitating real-time adjustment of the light source's incident angle. Based on the substrate surface roughness measurement data, the system automatically corrects the incident angle. Specifically, as the surface roughness increases, the system reduces the angle between the incident angle and the normal, aligning the light source's incident angle closer to the normal, thereby reducing reflection interference caused by surface roughness. This adjustment minimizes the impact of specular reflection on image quality, even on relatively rough surfaces.

[0071] The background difference method is used to subtract background data from the normal area of the target image to highlight the differential signal in the defective area. First, the system collects a light intensity distribution template from the normal area, representing the light intensity characteristics of the defect-free area. Subsequently, a dynamic weighted average algorithm is used to update the template to adapt to possible changes in the production process. By updating the background template in real time, the system can more accurately reflect the normal light intensity distribution under the current process state. Next, by performing differential processing on the light intensity data in the target image and the updated template, the light intensity of the normal area is subtracted, and the differential signal of the defective area is retained. This process can effectively remove background interference, make the signal in the defective area more obvious, and thus improve the sensitivity and accuracy of defect detection.

[0072] In one possible implementation, the feature vectors of the dielectric anomaly frequency band are first normalized by frequency band sensitivity to obtain a standardized feature representation. Frequency band sensitivity represents the responsiveness of different frequency bands to ceramic material defect detection. By normalizing the dielectric feature vectors based on the weights of the frequency band sensitivities, the network ensures that the features of different frequency bands are appropriately weighted, avoiding over- or under-influence of certain frequency band features. The normalized dielectric features serve as the key value in the attention mechanism.

[0073] Extracting local binary pattern texture features from defect images involves describing the texture of local regions in the image to obtain visual features that reflect the defect information. The image's texture features undergo dimensionality reduction processing in the convolutional layer, generating a more compact and effective feature representation that serves as the query value in the attention mechanism. This method extracts the image's visual features from a high-dimensional space and converts them into a low-dimensional representation, improving computational efficiency and recognition performance.

[0074] Next, a multi-head attention mechanism is used to calculate the cross-modal weight matrix between dielectric frequency band features and image texture features. The core idea of the attention mechanism is to calculate the importance weights for each pair of feature channels based on the relationship between the input query and key values. These weights represent the contribution of each feature to defect recognition in different frequency bands and image textures. This multi-head attention mechanism allows the system to simultaneously focus on multiple feature subspaces and automatically suppress feature channels unrelated to the current process deviation type during the learning process, ensuring the accuracy and effectiveness of recognition results.

[0075] During the training of the generative adversarial network, process deviation type labels are used as constraints for the generator. Based on these labels, the generator outputs synthetic defect morphologies. The generated samples are also affected by process parameter perturbations, allowing them to simulate defect morphologies under various process deviation conditions. The discriminator receives both real defect samples and synthetic samples generated by the generator. Through adversarial training, the generator is optimized to produce more realistic defect samples. This allows the generator to learn the distribution of defect morphologies under different process conditions, improving the system's ability to identify defects under complex process deviations.

[0076] In one possible implementation, when the system detects a new defect type, it first overlays a simulated defect texture onto a normal sample based on its associated process deviation type label. This simulated defect texture is generated by simulating the stress distribution caused by process parameter deviations (such as temperature, pressure, and material composition) using finite element analysis. Changes in stress distribution can cause cracks, holes, or other types of defects, and the morphological characteristics of these defects (such as crack width, depth, and shape) can be accurately predicted and simulated using finite element analysis.

[0077] Finite element analysis (FEA) is a computational method used to simulate stress distribution and deformation within a material. By inputting process parameters (such as temperature variations and applied pressure), the microstructure of the ceramic material is simulated to predict the effects of process deviations on the surface of the material, including cracks or holes. By simulating stress concentration and fracture processes, the geometry of defects can be accurately generated, including the topology of cracks, the distribution of holes, and other potential defect types.

[0078] After generating the simulated defect images, they need to be spatially aligned with the real dielectric scan data. This step ensures that the defect features in the simulated images match the actual process data, thereby forming a highly reliable training dataset. By spatially aligning the simulated defect images with the real scan data, the model can obtain more accurate training samples, thereby improving the model's ability to identify real process deviations. At this point, the augmented training dataset will include defect images under a variety of process conditions and provide more diverse training data, thereby improving the model's generalization and adaptability.

[0079] The augmented training dataset generated through the above steps is used to train the generative adversarial network. The generator learns to generate simulated defect images from normal samples and process deviation labels, while the discriminator is responsible for determining the differences between the generated defect images and real images. Through adversarial training, the generator is continuously optimized to produce more realistic defect images. The discriminator learns to distinguish between real and generated defect images, further improving the system's robustness in practical applications.

[0080] In one possible implementation, a coupled thermal-mechanical simulation model of the HTCC sintering process must first be constructed. This coupled thermal-mechanical model accounts for the effects of temperature and stress changes on the material. During the simulation, the current process parameters (such as temperature and pressure) and the combination of process parameters to be adjusted (such as sintering time, cooling rate, and pressure distribution) are input into the model. The model's goal is to comprehensively predict the temperature and stress fields of the material, as well as their impact on the ceramic material, under different process parameters during the sintering process.

[0081] During the sintering process of HTCC ceramic materials, thermal stress distribution occurs due to differences in the coefficient of expansion of different layers of material. This step uses a digital twin model to simulate the impact of these differences in coefficient of expansion and further calculate the stress distribution between material layers. These thermal stresses can cause defects such as cracks and holes. Therefore, this model can accurately predict the probability of defects in different areas and their distribution.

[0082] The results of the coupled thermal-mechanical simulation model predict which areas are likely to experience defects during sintering. By simulating differences in thermal stress and expansion coefficients, the model can infer the probability of defect generation and their distribution. This prediction, based on the material's physical properties and current process conditions, provides a scientific basis for subsequent parameter adjustments.

[0083] If the simulated defect rate reduction doesn't meet expectations, the system automatically adjusts parameters and reruns the simulation. By continuously adjusting process parameters, it simulates the changing trends in defect generation under different parameter combinations until the preset verification criteria are met. This iterative process ensures that process parameters can be precisely controlled in actual production, significantly reducing the occurrence of defects.

[0084] To optimize the adjustment process, the preset number of occurrences is closely tied to the defect severity level. By setting different defect severity thresholds, the higher the defect severity, the fewer consecutive occurrences are required to adjust the corresponding parameters. This strategy allows for more precise control over the adjustment response speed, ensuring that severe defects are effectively addressed in the shortest possible time, thereby preventing high-risk defects in the production process.

[0085] In one possible implementation, data is first collected from known defect samples, measuring the rate of change of the dielectric constant at different frequency bands and correlating this with the defect size. Specifically, the dielectric constant reflects the material's response characteristics at different electromagnetic frequencies, and the size and shape of the defect directly affect the rate of change of the dielectric constant. By experimentally measuring the dielectric constant changes at these frequency bands, a correlation coefficient can be obtained between each frequency band and the defect size. This correlation coefficient reveals the sensitivity of different frequency bands in identifying different defect types, providing basic data for subsequent analysis.

[0086] Based on the measured correlation coefficients, principal component analysis (PCA) is used to reduce the dimensionality of the dielectric constant variations in each frequency band. PCA aims to extract the principal components that maximize information retention through linear transformation, thereby selecting the frequency band combinations that contribute most to defect classification. PCA not only reduces the impact of redundant data but also helps identify the most critical features in multi-band data, thereby improving the accuracy of subsequent defect identification.

[0087] Based on the correlation coefficients obtained in the previous step and the weights from the principal component analysis, a comprehensive sensitivity score is calculated for each frequency band. This score reflects the importance of each frequency band in the defect identification process; higher scores indicate greater sensitivity to defects. To ensure that frequency bands with high sensitivity to defect types are prioritized during scanning, the scanning priority of frequency bands is adjusted based on this score. This allows the system to prioritize frequency bands that contribute significantly to defect classification and are highly sensitive, thereby improving the efficiency and accuracy of defect detection.

[0088] In one possible implementation, the substrate's surface roughness needs to be measured before automated HTCC ceramic defect detection. This step uses laser triangulation to measure the substrate's surface profile. Laser triangulation calculates the surface profile by emitting a laser beam onto the substrate surface and detecting the angular variation of the reflected beam. This highly accurate, non-contact method can accurately capture even subtle surface fluctuations, making it suitable for high-precision industrial inspection.

[0089] Surface profile data obtained through laser triangulation includes peak-to-valley height differences and root mean square (RMS) values. The peak-to-valley height difference indicates the vertical distance between the highest and lowest points on the surface, while the RMS value measures the overall fluctuation in surface roughness. Based on this data, the surface roughness grade can be calculated. For example, smaller peak-to-valley height differences and RMS values correspond to smoother surfaces, while larger values indicate a rougher surface. This calculation method typically adheres to specific industry standards to generate a surface roughness grade, providing critical information for subsequent scanning and defect identification.

[0090] Adjust the correction rules for the incident angle based on the surface roughness level. Specifically, for each increase in the roughness level, the angle between the incident angle and the normal will decrease by a preset angle step. This adjustment is based on the effect of the rough surface on the dielectric constant scan. When the surface roughness is high, adjusting the incident angle can help improve the quality of the scanning signal, because a rougher surface may cause signal scattering or uneven reflection, which in turn affects the recognition accuracy. By adjusting the incident angle, the scan can more accurately reflect the true characteristics of the ceramic surface, avoiding errors caused by surface roughness.

[0091] In one possible implementation, during the sintering process of HTCC ceramics, temperature gradients and interlaminar stresses are important factors affecting ceramic quality and defect formation. Therefore, these data need to be regularly collected during the actual sintering process. Temperature gradients reflect differences in thermal conductivity at different locations within the ceramic, while interlaminar stresses represent the stress distribution within different layers within the ceramic. These data can effectively describe the thermal-mechanical coupling characteristics of the sintering process and serve as the basis for calibrating simulation models.

[0092] The collected actual temperature gradient and interlaminar stress data are compared with the simulation model output. This comparison allows us to assess the accuracy of the simulation model and its consistency with the actual situation. If the simulation results differ from the actual situation, further adjustments to the model parameters are necessary to improve the accuracy of the simulation results.

[0093] When the error between the simulation results and the actual data exceeds the preset allowable range, the simulation model needs to be optimized using the backpropagation algorithm. This is a commonly used optimization method that calculates the relationship between the error and the material parameters and continuously adjusts the material parameters in the simulation model to make the simulation results more similar to the temperature gradient and interlaminar stress data during the actual sintering process.

[0094] When updating the simulation model, the key material parameters adjusted include the thermal conductivity, thermal expansion coefficient, and viscoelastic modulus of the HTCC green body.

[0095] Specifically: Thermal conductivity: affects the efficiency of heat conduction and is a key parameter that describes the thermal conductivity characteristics of materials.

[0096] Thermal expansion coefficient: determines the volume change of a material when heated and is an important factor in thermal-mechanical coupling.

[0097] Viscoelastic modulus: It characterizes the deformation characteristics of materials under stress and directly affects the deformation and stress distribution of ceramics.

[0098] By accurately calibrating these material parameters, the simulation model can be made more consistent with the actual thermal-mechanical behavior of the sintering process, thereby improving the model's predictive ability.

[0099] The calibration method for the thermal-mechanical coupling simulation model collects temperature gradient and interlaminar stress data from the actual sintering process and adjusts material parameters using a backpropagation algorithm to ensure consistency between the simulation model and actual conditions. This method not only improves the accuracy and predictive power of the simulation model but also provides more accurate theoretical support for the automatic identification of HTCC ceramic defects, thereby improving defect recognition rates and product quality control during industrial production.

[0100] The present invention encompasses any alternatives, modifications, equivalents, and solutions that fall within the spirit and scope of the present invention. To provide a thorough understanding of the present invention, specific details are described in detail below in connection with the preferred embodiments of the present invention, but those skilled in the art will be able to fully understand the present invention without these detailed descriptions. Furthermore, to avoid unnecessary confusion regarding the essence of the present invention, well-known methods, processes, procedures, components, and circuits have not been described in detail.

[0101] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. An automatic identification method for HTCC ceramic defects based on industrial model assistance, characterized in that: The following steps are involved: Step 1: Real-time collection of HTCC production process parameters including sintering temperature gradient, lamination pressure, and green body moisture content. A spatiotemporal baseline model of the process parameters is constructed using a long-short-term memory network. The deviation between the current parameters and the baseline model is dynamically calculated. When the deviation exceeds a dynamic threshold adaptively set based on historical stability data of the production line, a process deviation type label is generated. Step 2: Perform a non-contact dielectric property scan on the sintered HTCC substrate on the production line conveyor. Dynamically select the dielectric scanning frequency band based on the process deviation type label in step 1. Obtain the dielectric constant distribution map using the resonant cavity perturbation method. Calculate the local variance and mark the coordinates of the dielectric anomaly area. Map these coordinates to the optical detection coordinate system. Step 3: Based on the coordinates of the dielectric anomaly area, control the circularly polarized light source to illuminate the target area at an incident angle that suppresses reflection from the metallization layer. Dynamically match the near-infrared imaging wavelength based on the dielectric scanning frequency band, and subtract the light intensity of the normal area through the background difference method to generate a high-contrast image of the local defect. Step 4: Input the dielectric anomaly frequency band feature vector and the defect image texture features from step 3 into the cross-modal attention network to generate a fused feature map. Activate the corresponding defect classification branch based on the process deviation type label, and use a generative adversarial network to separate the defect features from the process background texture, outputting the defect type and severity level. Step 5: Establish an association mapping library between defect types and process deviation types. When the same defect type appears a preset number of times consecutively, automatically adjust the sintering furnace temperature control parameters, verify the change in defect rate after parameter adjustment through the digital twin model, update the dynamic threshold of process deviation, and form a closed-loop feedback loop.

2. The method for automatically identifying HTCC ceramic defects based on industrial model assistance according to claim 1, characterized in that: The method for determining the adaptive dynamic threshold based on the historical stability data of the production line in step 1 includes: Collect temperature gradient data, lamination pressure time series data, and green body moisture content detection data in each temperature zone during the continuous production cycle of the production line to build a historical database of process parameters; Divide historical data by production batches, calculate the standard deviation and mean shift of parameters within each batch, and generate stability evaluation indicators; Based on the stability evaluation index corresponding to the current production batch, a sliding window algorithm is used to dynamically adjust the deviation judgment threshold: when the stability evaluation index is lower than the preset critical value, the tolerance range of the deviation judgment threshold is expanded; when the index is higher than the critical value, the threshold range is narrowed; The generation of the process deviation type label includes: matching the current parameter deviation pattern with a predefined process abnormality pattern library, wherein the pattern library is established by clustering historical abnormality data, and each pattern contains a deviation weight combination of temperature, pressure, and moisture content.

3. The method for automatically identifying HTCC ceramic defects based on industrial model assistance according to claim 1, characterized in that: The method for dynamically selecting the dielectric scanning frequency band according to the process deviation type label in step 2 includes: Construct a dielectric frequency band-defect type sensitivity mapping table: experimentally measure the dielectric parameter change rate of various defects at different frequency bands and calculate the frequency band's response intensity to the defect; Based on the process deviation type label in step 1, select the top N frequency bands with the highest response intensity from the mapping table as the scanning frequency bands. N is dynamically adjusted based on the real-time requirements of the production line inspection. The mapping method of the dielectric abnormal area coordinates includes: setting a reference positioning mark on the conveyor belt, obtaining the real-time position of the substrate through a laser rangefinder, establishing an affine transformation relationship between the dielectric scanning coordinate system and the optical imaging coordinate system, and achieving submillimeter spatial alignment.

4. The method for automatically identifying HTCC ceramic defects based on industrial model assistance according to claim 1, characterized in that: The method for adjusting the incident angle of the circularly polarized light source to suppress reflection of the metallized layer in step 3 includes: Collect reflectivity data of the metallization layer on the surface of the HTCC substrate, train the light source angle optimization model through historical imaging samples, and output the optimal incident angle range for suppressing specular reflection; The incident angle is corrected in real time based on the surface roughness detection data of the current substrate. The angle between the incident angle and the normal is reduced when the roughness is higher. The implementation of the background difference method includes: collecting the light intensity distribution template of the normal area, updating the template through a dynamic weighted average algorithm, and deducting the template data from the target image to retain the difference signal of the defect area.

5. The method for automatically identifying HTCC ceramic defects based on industrial model assistance according to claim 1, characterized in that: The method for constructing the cross-modal attention network by inputting the dielectric anomaly frequency band feature vector and the defect image texture feature into the cross-modal attention network in step 4 includes: Normalize the dielectric frequency band feature vector according to the frequency band sensitivity weight and use it as the key value of the attention mechanism; Extract the local binary pattern texture features of the defect image and use them as query values after dimension reduction through the convolution layer; A multi-head attention mechanism is used to calculate the cross-modal weight matrix and suppress feature channels that are irrelevant to the current process deviation type. The constraints of the adversarial generative network include: limiting the defect morphology distribution output by the generator according to the process deviation type label, and the discriminator receiving real defect samples and synthetic samples generated by process parameter perturbations for adversarial training.

6. The method for automatically identifying HTCC ceramic defects based on industrial model assistance according to claim 5, characterized in that: The training data generation method for separating defect features and process background textures using a generative adversarial network in step 4 includes: When a new type of defect is detected, a simulated defect texture is superimposed on the normal sample according to its associated process deviation type label; The method for generating the simulated defect texture includes: simulating the stress distribution caused by the deviation of the process parameters through finite element analysis to generate the morphological characteristics of the crack or hole; The simulated defect images are spatially aligned with the real dielectric scanning data to construct an augmented training dataset.

7. The method for automatically identifying HTCC ceramic defects based on industrial model assistance according to claim 1, characterized in that: The method for verifying the change in defect rate after parameter adjustment by using the digital twin model in step 5 includes: Build a thermal-mechanical coupling simulation model for the HTCC sintering process, input the current process parameters and the parameter combination to be adjusted; Simulate the expansion coefficient difference and thermal stress distribution between layers during sintering to predict the probability and distribution of defects; When the reduction in the simulated defect rate does not meet expectations, iteratively adjust the parameters and re-simulate until the verification conditions are met; The method for determining the preset number of times includes: setting different thresholds according to the severity level of the defect, and the higher the severity, the fewer consecutive occurrences are required to trigger parameter adjustment.

8. The method for automatically identifying HTCC ceramic defects based on industrial model assistance according to claim 3, characterized in that: The method for establishing the dielectric frequency band-defect type sensitivity mapping table includes: In samples with known defects, the correlation coefficient between the dielectric constant change rate and the defect size in each frequency band is measured; Use principal component analysis to screen the frequency band combination that contributes most to defect classification; The comprehensive sensitivity score of the frequency band is calculated based on the correlation coefficient and the principal component weight. The higher the score, the higher the scanning priority.

9. The method for automatically identifying HTCC ceramic defects based on industrial model assistance according to claim 4, characterized in that: The method for acquiring the surface roughness detection data comprises: Before dielectric scanning, the substrate surface roughness profile was measured by laser triangulation; Calculate the surface roughness grade based on the peak-to-valley height difference and root mean square value of the profile data; The incident angle correction rule is: for each increase in the roughness level, the angle between the incident angle and the normal line decreases by a preset angle step.

10. The method for automatically identifying HTCC ceramic defects based on industrial model assistance according to claim 7, characterized in that: The calibration method of the thermal-mechanical coupling simulation model includes: Regularly collect temperature gradient data and interlaminar stress detection data during the actual sintering process and compare them with the simulation results; When the error exceeds the allowable range, the material parameter boundary conditions of the simulation model are updated through the back propagation algorithm; The material parameters include thermal conductivity, thermal expansion coefficient and viscoelastic modulus of the HTCC green body.

Images