Method and device for detecting defects of thin-wall layer coated with solder paste

Abstract

The invention relates to a method and a device for detecting defects of a thin-wall layer coated with solder paste. The method comprises the following steps: performing dual-wavelength polarized light scanning on thin-wall layer solder paste on a PCB (Printed Circuit Board) to obtain a polarization spectrum data set; performing pulse laser irradiation and ultrasonic signal acquisition on the thin-wall layer solder paste to obtain a thermo-acoustic coupling response data set; performing fluorescence lifetime imaging scanning on the thin-wall layer solder paste mixed with the quantum dots in advance to obtain a fluorescence parameter data set; and inputting the polarization spectrum data set, the thermal-acoustic coupling response data set and the fluorescence parameter data set into an attention guidance deformation convolutional network for feature extraction and fusion, and outputting a comprehensive detection report. According to the invention, the detection capability of different depth areas is enhanced, the accurate characterization of the internal acoustic characteristics of the thin-wall layer is realized, and the defect detection precision of the thin-wall layer is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of defect detection, in particular to a defect detection method and device for thin-wall layer of tin paste coating. BACKGROUND

[0002] With the acceleration of the miniaturization trend of electronic products, the tin paste printing process of miniaturized PCB in the electronic manufacturing industry is facing severe challenges, especially in the quality control of thin-wall layer structure area. The traditional optical detection method is limited by imaging resolution and penetration depth, and it is difficult to effectively identify the micro-defects inside the thin-wall structure, especially in low-contrast areas and hidden positions, there is obvious detection blind area. Although X-ray tomography can provide certain depth information, the detection rate of micro-defects is insufficient, and the equipment cost is high and the detection efficiency is low, which is difficult to meet the online detection demand.

[0003] The wide application of thin-wall micro-channel structure and ultra-thin-wall micro-cavity structure makes the tin paste coating quality become a key factor affecting product reliability. The core difficulty of existing detection technology is that it cannot simultaneously meet the requirements of detection depth, detection accuracy and processing speed, especially for the detection ability of hidden defects such as bubbles, micro-cracks and composition segregation in thin-wall area, which leads to a large number of potential defects cannot be discovered in time in the manufacturing link. SUMMARY

[0004] The main purpose of the present application is to provide a defect detection method and device for thin-wall layer of tin paste coating, which enhances the detection ability of different depth areas, realizes the accurate characterization of the acoustic characteristics inside the thin-wall layer, and improves the accuracy of thin-wall layer defect detection.

[0005] To achieve the above purpose, the present application provides a defect detection method for thin-wall layer of tin paste coating, comprising: performing double-wavelength polarized light scanning on the thin-wall layer tin paste on the PCB to obtain a polarized spectrum data set; performing pulsed laser irradiation and ultrasonic signal acquisition on the thin-wall layer tin paste to obtain a thermal-acoustic coupling response data set; performing fluorescence lifetime imaging scanning on the thin-wall layer tin paste mixed with quantum dots in advance to obtain a fluorescence parameter data set; inputting the polarized spectrum data set, the thermal-acoustic coupling response data set and the fluorescence parameter data set into an attention-guided deformable convolution network for feature extraction and fusion, and outputting a comprehensive detection report.

[0006] The present application also provides a defect detection device for thin-wall layer of tin paste coating, comprising: a scanning module for performing double-wavelength polarized light scanning on the thin-wall layer tin paste on the PCB to obtain a polarized spectrum data set; The collection module is used for pulse laser irradiation and ultrasonic signal collection on the thin-wall layer tin paste, so as to obtain a thermal-acoustic coupling response data set; The imaging module is used for fluorescence lifetime imaging scanning on the thin-wall layer tin paste mixed with quantum dots in advance, so as to obtain a fluorescence parameter data set; The output module is used for inputting the polarized spectrum data set, the thermal-acoustic coupling response data set and the fluorescence parameter data set into an attention-guided deformable convolution network for feature extraction and fusion, and outputting a comprehensive detection report.

[0007] In conclusion, the technical scheme provided by the present application realizes the all-around detection capability of the thin-wall layer tin paste by integrating the dual-wavelength polarized light scanning, thermal-acoustic coupling response detection and quantum dot fluorescence lifetime imaging multiple technical means, and significantly breaks through the technical bottleneck of the traditional single detection means. Through the synergistic effect of the 445nm and 785nm dual-wavelength polarized light sources, the detection capability of different depth regions is enhanced; by means of the ultrasonic signal analysis generated by the thermal elastic effect, the accurate characterization of the acoustic characteristics inside the thin-wall layer is realized; by using the fluorescence characteristic change of the CdSe / ZnS quantum dots as a tracer, the sensitivity to the micro defects is effectively improved; by using the attention-guided deformable convolution network for feature fusion, the detection problem of the irregular thin-wall boundary and the low-contrast region is solved; by means of the multi-threshold self-adjusting segmentation and depth estimation algorithm, the three-dimensional structure reconstruction of the defects is realized; by combining the multi-task learning analysis of the tin paste metal characteristic library, the complete quality control process from defect detection to grading warning is established, the defect warning time is advanced to the printing process, and the system processing speed meets the online detection demand, so that the present application can be seamlessly integrated into the existing electronic manufacturing production line, and reliable quality guarantee is provided for high-end electronic product manufacturing. BRIEF DESCRIPTION OF DRAWINGS

[0008] Figure 1 is a defect detection method step schematic diagram of a tin paste coated thin-wall layer in an embodiment of the present application; Figure 2 is a structural block diagram of a defect detection device of a tin paste coated thin-wall layer in an embodiment of the present application.

[0009] The implementation, functional characteristics and advantages of the present application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0010] In order to make the purpose, technical scheme and advantages of the present application more clear, the present application will be further described in detail below by combining the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and are not used to limit the present application.

[0011] Reference Figure 1The embodiment provides a thin-wall layer defect detection method for tin paste coating, and comprises the following steps: S1, performing double-wavelength polarized light scanning on the tin paste of the thin-wall layer on the PCB board to obtain a polarized spectrum data set; The blue laser and the near-infrared laser are combined by a dichroic beam splitter to form a dual-wavelength beam containing two specific wavelengths, where the wavelength of the blue laser is 445 nm and the wavelength of the near-infrared laser is 785 nm. The dichroic beam splitter uses the transmission and reflection characteristics of specific wavebands to make the two different wavelengths of laser propagate along the same optical path, ensuring the spatial consistency of the beam and maintaining the stability of the beam quality. The dual-wavelength beam enters a polarization beam splitter for polarization separation. Since the polarization beam splitter can separate the polarization components of the incident light according to the orthogonal relationship of the polarization state, this step separates the originally mixed beam into mutually orthogonal polarized light. The orthogonal polarized light passes through a 1 / 4 wave plate for polarization state conversion. The 1 / 4 wave plate converts linearly polarized light into circularly polarized light, so that the light of different polarization directions forms components with equal amplitude and a phase difference of 90° after passing through the wave plate, thereby generating circularly polarized light. The circularly polarized light carries rich polarization information in the reflected light when it is incident on a non-uniform surface, which helps to extract the microstructure features of the thin-wall layer surface of the solder paste. The circularly polarized light passes through a collimating and focusing system composed of a collimating lens and a focusing lens. The collimating lens adjusts the divergence angle of the beam to make it a collimated beam, and the focusing lens focuses the collimated beam to the surface of the thin-wall layer of the solder paste, forming a small spot with high energy density. The diameter of the spot is precisely controlled to be 25±2 μm to ensure that the microstructure features of the solder paste surface and internal can be detected, while avoiding the reduction of spatial resolution due to the oversize spot. When the circularly polarized light irradiates the surface of the thin-wall layer of the solder paste, part of the light is absorbed by the solder paste, and part of the light is scattered and reflected. The reflected light carries the topographic information of the solder paste surface and the polarization state changes caused by the internal structure. These reflected lights then return along the coaxial optical path and are received by a polarization analyzer array arranged in the optical path. The polarization analyzer array is composed of 16 independent phase detection units, and the detection angle of each unit is separated by 22.5°, covering the complete 0-360° polarization state range. This design enables the system to simultaneously collect reflected signals of multiple different polarization angles, thereby constructing the complete polarization feature information of the thin-wall layer of the solder paste. The polarization state change information of the dual-wavelength channel is simultaneously collected by a photomultiplier tube to record the polarization response of the solder paste under two different wavelengths.In order to improve the accuracy and stability of the data, the collected polarization state change information is preprocessed, including dark background subtraction, non-uniformity correction and polarization degree calculation. The purpose of dark background subtraction is to remove the influence of environmental light and system noise, and to ensure that the final data only contains effective polarization signals from the solder paste. The non-uniformity correction mainly compensates for the non-uniformity of the optical system and the detector, so that the polarization information at different positions has consistent response standards, thereby improving the reliability of the data. Then, the anisotropic characteristics of the solder paste surface are quantitatively analyzed by calculating the polarization degree. The calculation of the polarization degree is based on Stokes parameters, which can effectively represent the scattering characteristics and internal microstructure distribution of the solder paste surface. Finally, a standardized polarized spectrum data set is obtained, which contains the spatial coordinate information, wavelength information and polarization state change information of the thin-wall layer of the solder paste.

[0012] S2, pulse laser irradiation and ultrasonic signal acquisition are performed on the thin-wall layer of the solder paste to obtain a thermal-acoustic coupling response data set; Specifically, a Q-switched pulsed laser is used, with the output wavelength set to 1064 nm, the pulse width to 15 ns, the repetition frequency to 10 kHz, and the single-pulse energy density controlled to 5-8 mJ / cm2, to ensure that the laser irradiation can generate sufficient thermal-elastic effect without damaging the solder paste material, thereby exciting ultrasonic waves on the surface and inside of the solder paste. The laser beam is guided to the surface of the thin-wall layer of solder paste through an optical scanning system, and the solder paste is irradiated point by point in a precise scanning path. The laser focusing diameter of each scanning point is set to 30 pm, and the scanning pitch is 15 pm, to ensure a 50% overlap rate so that no small area is missed during scanning, while maintaining the continuity and uniformity of the thermal-acoustic excitation signal. When the pulsed laser beam irradiates the surface of the solder paste, the short-time deposition of laser energy will cause the solder paste material to be locally heated and generate a transient temperature gradient, thereby triggering the thermal-elastic effect and generating ultrasonic signals inside the solder paste. These ultrasonic waves propagate in the thin-wall layer of solder paste, and their propagation characteristics will be affected by internal defects such as bubbles, cracks, voids, and composition segregation, resulting in changes in the propagation path, phase, amplitude, and spectral characteristics of the ultrasonic signal. Therefore, while laser scanning is performed, a high-sensitivity piezoelectric sensor array is used to receive and collect the ultrasonic signals generated on the surface of the solder paste. The sensor array is composed of 64 independent high-frequency piezoelectric elements, can cover a wide frequency range of 0.5-15 MHz, and has a high sensitivity of -60 dB re V / pPa, to ensure effective capture of weak ultrasonic signals. During signal acquisition, each piezoelectric element synchronously records the received ultrasonic waveform, which is stored in real time by a high-speed data acquisition system with a sampling rate of 100 MS / s, to ensure that the time resolution of the signal is high enough to accurately analyze the ultrasonic propagation characteristics inside the solder paste. To improve the signal-to-noise ratio of the ultrasonic signal, the original ultrasonic signal is subjected to signal enhancement and filtering processing. The received original ultrasonic signal is amplified by a low-noise preamplifier with a gain of 40 dB, to ensure that the signal has sufficient dynamic range before entering the subsequent filtering link, while reducing noise interference. The amplified signal is filtered by a bandpass filter with a passband range of 1-10 MHz, to effectively remove low-frequency noise and high-frequency interference components, and only retain the ultrasonic signal components highly related to the structural characteristics of the solder paste. At the same time, to capture the thermal response characteristics of the surface and internal structure of the solder paste, an infrared thermal imager is used to synchronously collect thermal imaging of the surface of the thin-wall layer of solder paste, with a temperature resolution of 0.05°C and a spatial resolution of 50 pm, to accurately record the transient temperature distribution on the surface of the solder paste and generate temperature field data varying with time wherein is a spatial coordinate, For the time dimension, the temperature distribution map reflects the temperature change trend of the local area of the solder paste after laser irradiation. The processed ultrasonic signal and transient temperature distribution map are registered, the position information of the laser scanning system is used to align the coordinates of the ultrasonic signal data and the temperature distribution data, so that they are matched at the same spatial position. Through timestamp matching technology, the time information of the infrared thermal imager is synchronized and corrected with the time acquisition sequence of the ultrasonic signal, to ensure that the ultrasonic waveform and temperature data collected under the same laser pulse action maintain time consistency, thereby constructing a thermal-acoustic coupling response data set .

[0013] S3, fluorescent lifetime imaging scanning is performed on the thin wall layer solder paste with quantum dots mixed in advance, and a fluorescent parameter data set is obtained; It should be noted that the amount of CdSe / ZnS core-shell structure quantum dots is uniformly incorporated in the solder paste, the core diameter of the quantum dots is controlled to be 4-7 nm, and the surface is modified with carboxyl functional groups to ensure good compatibility with the solder paste matrix, and the mass fraction of the quantum dots is controlled to be 0.01%-0.05%, which ensures sufficient fluorescence signal intensity and does not significantly affect the physical properties of the solder paste material. After the solder paste is coated on the thin wall layer of the PCB board and solidified, a picosecond pulse laser is used to excite and irradiate the solder paste, the center wavelength of the laser is set to 405 nm, the pulse width is controlled to be 80 ps, and the repetition frequency is 80 MHz to ensure high time resolution of the fluorescence lifetime measurement, and the scanning path of the laser on the surface of the solder paste is accurately controlled through a scanning galvanometer system, so that the fluorescence signal of each point can be fully excited to obtain complete spatial fluorescence information. Under laser irradiation, the quantum dots in the solder paste absorb laser energy and transition to the excited state, then transition back to the ground state by radiation and release fluorescence signals, and the intensity, emission peak position and lifetime of the fluorescence signals are affected by the surrounding environment, so the microstructure and defects in the solder paste are analyzed by detecting these parameters. In order to realize accurate measurement of the fluorescence signal of the quantum dots, a dichroic filter is used to separate the fluorescence signal by wavelength, the cutoff wavelength of the filter is set to 450 nm to ensure that the excitation light can be effectively filtered out and only the fluorescence signal of the quantum dots can pass through, and since the fluorescence emission peak of the quantum dots is located at 570-590 nm, the subsequent detector can efficiently collect the fluorescence signal in this wavelength band. The filtered fluorescence signal is received and converted by a high-sensitivity photomultiplier tube, the quantum efficiency of the PMT is maintained at more than 40% in the range of 450-650 nm, which can ensure efficient detection of weak fluorescence signals, and the internal high-gain amplification module converts the optical signal into an electrical signal to generate a fluorescence electrical signal for subsequent time sequence analysis. After obtaining the fluorescence electrical signal, a time-correlated single-photon counting technique is used to collect the signal with high time resolution, which constructs a complete fluorescence decay curve by counting the distribution of photon arrival times, and since the time resolution of the TCSPC technique reaches 25 ps, it can accurately depict the fluorescence lifetime characteristics of the quantum dots and effectively distinguish the fluorescence decay behavior in different environments. During data acquisition, each sampling point accumulates at least 1000 single-photon events to ensure the statistical stability of the signal, and the fluorescence decay curve of each point is normalized to eliminate the non-uniformity of the system response. The fluorescence decay curve is fitted by a three-exponential model through least squares method, and the goodness of fit is required to be greater than 0.995 to ensure the calculation accuracy of the fluorescence lifetime. When the quantum dots are in the normal solder paste matrix, the main fluorescence lifetime component about 15-20 ns, while when the quantum dots are in a defect region (such as a bubble, a crack or an oxidation area), due to the change of the local environment, the fluorescence lifetime is shortened to 5-10 ns, accompanied by a blue shift (5-15 nm) or a red shift (10-20 nm) of the fluorescence emission peak, therefore, by analyzing the spatial distribution of the fluorescence lifetime parameters, potential defects in the thin wall layer of the solder paste can be effectively identified. The fluorescence lifetime parameters, the fluorescence intensity and the emission peak position information are subjected to spatial filtering and denoising processing, the fluorescence intensity map is smoothed by using Gaussian filtering to reduce the influence of random noise, at the same time, the fluorescence lifetime parameters are preliminarily classified by using an adaptive threshold method, a fluorescence anomaly index is defined, and the abnormal area is marked. Through principal component analysis, the three characteristics of the fluorescence lifetime, the fluorescence intensity and the emission peak position are reduced in dimension, so as to extract the main characteristic components, and the local neighborhood statistical method is used to accurately detect the boundary of the defect area, and finally the fluorescence parameter data set is obtained.

[0014] S4, inputting the polarization spectrum data set, the thermo-acoustic coupling response data set and the fluorescence parameter data set into an attention-guided deformation convolution network for feature extraction and fusion, and outputting a comprehensive detection report.

[0015] Specifically, each dataset was pre-processed to ensure that they could be calculated and analyzed in the same feature space. The polarized spectral matrix P(x, y, l, q) in the polarized spectral dataset was converted into a two-dimensional image matrix through wavelength channel combination and polarization state mapping processing. This process normalized the polarized light data at different wavelengths and reduced the dimensionality through principal component analysis to extract the most representative polarization information. Meanwhile, the pseudo-color mapping was constructed using the degree of polarization and the two-wavelength polarization contrast to convert the high-dimensional spectral data into a two-dimensional feature image that could be used for deep learning processing. At the same time, the acoustic impedance distribution Z(x, y, z) and the transient temperature distribution T(x, y, t) in the thermo-acoustic coupling response dataset needed to be projected into two-dimensional features. This process used a spatiotemporal mapping algorithm to perform a deep weighted average of the acoustic impedance and a two-dimensional interpolation method to project the three-dimensional temperature data into a two-dimensional coordinate system matching the polarized spectral data, resulting in a unified scale of thermo-acoustic features. Meanwhile, the fluorescence lifetime, intensity, and emission peak position information in the fluorescence parameter dataset were reconstructed to form two-dimensional feature distribution maps. The fluorescence lifetime distribution map was calculated by fitting a three-exponential model and spatially interpolating the lifetime parameters in different regions. The fluorescence intensity and spectral shift information were denoised using Gaussian filtering and adaptive mean filtering to obtain high-quality fluorescence feature images. After completing the data conversion, the two-dimensional image matrices were spatially registered based on the affine transformation algorithm to ensure that they corresponded to the same physical region of the multi-modal image dataset. Key points in the polarized spectral image, thermo-acoustic feature image, and fluorescence feature image were extracted using the scale-invariant feature transform algorithm, and the corresponding feature descriptors were calculated. Then, the nearest neighbor matching method was used for preliminary alignment, and the random sample consensus algorithm was used to remove false matching points to improve registration accuracy. Next, an affine transformation matrix was used to globally transform all images to align them in the same spatial coordinate system, forming a registered multi-modal dataset. This multi-modal image dataset was input into the backbone feature extraction network of the attention-guided deformable convolutional network for depth separable convolution processing to obtain an initial feature map containing information about the thin-wall tin paste structure. The backbone network uses a five-layer depth separable convolution structure, with a 3x3 convolution kernel size and a step size of 1. Batch normalization and PReLU activation functions are used for normalization processing to ensure the stability of feature expression. The initial feature map is input into the polarization feature branch, acoustic feature branch, and fluorescence feature branch for parallel feature extraction. The polarization feature branch extracts the degree of polarization and two-wavelength difference information to analyze the optical anisotropy on the surface and inside the tin paste. The acoustic feature branch extracts the sound speed and spectral features to reflect the acoustic propagation characteristics inside the tin paste. The fluorescence feature branch extracts the fluorescence lifetime and spectral shift information to identify the microenvironment changes and defect region distribution of the tin paste material.After feature extraction, the weights of the three sets of channel feature maps are dynamically adjusted using a channel attention mechanism to enhance the complementarity between different physical features. This attention mechanism employs a SE module to calculate the feature response weights for each channel through global average pooling and rescales them through a fully connected layer. This dynamically distributes the weights of each feature channel, ensuring that the polarization, acoustic, and fluorescence features complement each other and improving the overall feature representation. Furthermore, the receptive field shape of the weighted feature maps is automatically adjusted using a deformable convolution module. This module learns pixel-level offsets to adapt the convolution kernel to the complex boundary structure of the solder paste thin layer, thereby improving the feature representation of the boundary region and reducing errors caused by irregular thin-wall morphology. After feature fusion, the feature map adapted to the thin-wall boundary undergoes dimensionality reduction to reduce computational complexity and highlight key features. This dimensionality reduction process uses 1×1 convolution for channel dimensionality reduction, combined with a max pooling operation to extract the most globally significant features, resulting in a comprehensive feature map containing comprehensive feature information of the thin-wall solder paste layer. Based on this comprehensive feature map, thin-walled solder paste defects are detected using a multi-threshold self-adjusting segmentation algorithm and a depth estimation algorithm to generate a defect mask and a three-dimensional defect model, respectively. The multi-threshold self-adjusting segmentation algorithm, combined with the multi-scale feature pyramid structure of a U-Net network, uses region growing and boundary-aware modules to accurately segment defect regions and adaptively adjust the segmentation threshold based on local statistical characteristics. The depth estimation algorithm combines the propagation time difference of ultrasonic signals and fluorescence lifetime distribution data to calculate the depth distribution of solder paste defects and generate a three-dimensional defect model. The defect mask and three-dimensional defect model are then combined with a solder paste metallurgical feature library for multi-task learning analysis to determine the defect type, severity, and potential impact. This analysis uses a deep neural network for feature matching and combines it with a priori knowledge base for defect classification. Ultimately, a comprehensive inspection report is generated, including a defect distribution map, a three-dimensional reconstructed model, and quantitative evaluation data.

[0016] In one example, a dual-wavelength polarized light scan is performed on a thin layer of solder paste on a PCB to obtain a polarization spectrum dataset, including: The lasers generated by the blue laser and the near-infrared laser are combined by a dichroic beam splitter to obtain a dual-wavelength beam, and the dual-wavelength beams are polarized and separated by a polarization beam splitter to obtain orthogonal polarized light; The orthogonal polarized light is converted into circularly polarized light through a quarter wave plate, and the circularly polarized light is focused through a collimating and focusing system to obtain a surface spot focused on the thin-walled solder paste layer; The polarization analyzer array is used to detect the phase of the light signal reflected from the surface of the thin-walled solder paste layer to obtain the polarization state change information of the dual-wavelength channel; The polarization state variation information of the dual-wavelength channel is dark background subtracted, non-uniformity corrected and polarization degree calculated to obtain a polarized spectrum data set.

[0017] In this example, a blue laser and a near-infrared laser with appropriate wavelengths are selected, respectively, wherein the wavelength of the blue laser is set to 445 nm, and the wavelength of the near-infrared laser is set to 785 nm, and the two wavelengths of laser are used to extract the surface topography information and internal structure characteristics of the thin wall layer of the tin paste, respectively. In order to realize the beam combination of the dual-wavelength beams, a dichroic beam splitter is used, and the transmission and reflection characteristics thereof are designed to have high transmittance for and high reflectivity for has high reflectivity. When the blue laser beam is incident from the transmission end of the beamsplitter, it can directly pass through the mirror surface, while the near-infrared laser beam is incident at an angle and is reflected on the surface of the beamsplitter, so that the two beams of light are spatially overlapped to form a stable dual-wavelength beam. In this process, in order to ensure uniform power distribution of the two laser beams, an optical power control module is used for power balancing adjustment, so that the energy ratio of the dual-wavelength beam after beam combination meets the experimental requirements, and a high-precision beam shaping system is used to optimize the beam quality, so as to reduce the influence of beam divergence on the subsequent detection accuracy. After the dual-wavelength beam is combined, it is polarized and separated by a polarizing beam splitter to obtain orthogonal polarized light. Since the surface and internal structure of the solder paste material will produce different responses to light of different polarization states, accurate control of polarized light can effectively improve the sensitivity and accuracy of detection. The polarizing beam splitter uses a polarizing beam splitter prism structure, and its working principle is based on the Brewster angle effect, that is, when the incident light passes through the polarizing beam splitter, the s-polarized component and the p-polarized component will be spatially separated, thereby forming two beams of orthogonal polarized light. In order to ensure the quality of the light beam after polarization splitting, a compensation optical element is added after the light beam passes through the beam splitter to reduce the polarization distortion caused by system errors and further improve the polarization purity of the light beam. The orthogonal polarized light passes through a 1 / 4 wave plate for polarization state conversion to generate circularly polarized light. In this process, the phase delay characteristic of the wave plate is crucial, which converts linearly polarized light into orthogonal components with equal amplitude and a phase difference of 90°, so that the synthesized light field becomes circularly polarized light, ensuring that the light beam has a uniform polarization state distribution when it is incident on the surface of the solder paste, thereby avoiding the influence of polarization changes caused by different incident angles on the detection results. After the light beam passes through the wave plate, it is collimated and focused by a collimating focusing system composed of a collimating lens and a high numerical aperture focusing lens. The collimating lens is used to reduce the beam divergence angle and improve the parallelism of the light beam, while the focusing lens is used to focus the collimated circularly polarized light onto the surface of the solder paste thin wall layer, so that the incident light spot diameter is controlled at d=25±2μm, to ensure accurate detection of a small area. When the circularly polarized light irradiates the surface of the solder paste, part of the light is absorbed by the solder paste material, and another part of the light is reflected on the surface. At the same time, due to the non-uniformity of the solder paste material, the polarization state of the reflected light will change, and these changes contain information about the optical properties and surface topography of the material. In order to accurately detect the reflected light, a polarized analyzer array is used, which is composed of independent polarization detection units, and the detection angle interval of each unit is set to , so that the entire array can cover the full to Polarization angle range. After the reflected light signal passes through the polarization analyzer, the polarization state change information of the dual-wavelength channel is synchronously collected by a high-sensitivity photomultiplier tube and converted into an electrical signal for subsequent data analysis and processing. In order to improve the accuracy of the data, the collected dual-wavelength polarization state change information is subjected to dark background subtraction, non-uniformity correction and polarization degree calculation. The purpose of dark background subtraction is to remove the influence of external environmental light and system noise, by recording the dark current signal of the system without laser irradiation, and subtracting it from the actual detection signal. Non-uniformity correction is to compensate for the non-uniformity of the optical system response, so that the polarization data at different positions have consistent response standards, thereby improving the measurement stability of the system. The polarization degree calculation is based on the calculation method of Stokes parameters to quantitatively characterize the polarization properties of the solder paste surface. The calculation of the polarization degree can reveal the anisotropy characteristics of the solder paste material, and then be used to identify the differences in surface morphology and internal structure of the material. Through the above steps, the polarization spectrum data set is obtained.

[0018] In one example, the thin-wall layer solder paste is subjected to pulsed laser irradiation and ultrasonic signal collection to obtain a thermal-acoustic coupling response data set, including: The laser beam generated by the Q-switched pulsed laser is scanned and irradiated on the thin-wall layer solder paste to obtain ultrasonic waves generated by thermal-elastic effect, and the ultrasonic waves generated on the surface of the thin-wall layer solder paste are received and collected by a piezoelectric sensor array to obtain original ultrasonic signals; The original ultrasonic signals are subjected to signal enhancement and filtering processing by a low-noise preamplifier and a band-pass filter to obtain processed ultrasonic signals, and the surface of the thin-wall layer solder paste is subjected to synchronous thermal imaging collection by an infrared thermal imager to obtain a transient temperature distribution map; The processed ultrasonic signals and the transient temperature distribution map are subjected to registration processing to obtain a thermal-acoustic coupling response data set.

[0019] In this example, a suitable Q-switched pulsed laser is selected, the center wavelength of which is set to nm, the pulse width ns, the repetition frequency kHz, and the single-pulse energy density control in 5-8 mJ / cm2, to ensure that the laser energy can effectively stimulate the thermoelastic ultrasonic wave without damaging the solder paste material. When the pulsed laser beam irradiates the surface of the solder paste, due to the laser energy deposited in the local area in a short time, it will cause the local area to produce a transient temperature gradient, thus triggering thermal expansion and stress wave in the solder paste material, forming an ultrasonic signal, the wavelength, frequency and propagation characteristics of the ultrasonic signal are affected by the thermophysical properties, elastic modulus and internal defects of the solder paste material, so the microstructure of the solder paste and the possible defects are identified by analyzing the characteristics of the ultrasonic signal. In order to detect the thermoelastic ultrasonic wave of the solder paste surface with high precision, a piezoelectric sensor array composed of high-frequency piezoelectric sensors is used, the working frequency range of the array is set to MHz, and it has high sensitivity dB re V / uPa, to ensure that it can accurately receive the ultrasonic signal induced by laser. The piezoelectric sensor array is arranged in the specific detection area of the solder paste thin wall layer, and is in contact with the detection surface through a coupling medium such as ultrasonic gel, to optimize the coupling efficiency of the acoustic signal. Because the internal structure of the solder paste is not uniform, the sound velocity distribution of the defect area will change locally, which will affect the propagation path of the ultrasonic wave, resulting in phase delay and amplitude attenuation of the ultrasonic signal, and these change information is used for precise positioning and quantitative analysis of the defect. After receiving the original ultrasonic signal of the solder paste surface, the signal is enhanced and filtered by a low-noise preamplifier and a band-pass filter to improve the signal-to-noise ratio. The original ultrasonic signal is enhanced by a low-noise preamplifier with a gain of dB, so that the signal amplitude is increased, and the interference of environmental noise is reduced. The amplified signal is filtered by a band-pass filter with a center frequency of MHz and a bandwidth of MHz, to remove low-frequency background noise and high-frequency interference signals, and finally obtain the processed ultrasonic signal . At the same time, in order to record the temperature change of the solder paste surface synchronously, a high-resolution infrared thermal imager is used for thermal imaging acquisition, the temperature resolution of the thermal imager is set to °C, the spatial resolution is set to μm, and the frame rate is set to Hz for real-time imaging. Because the defect area (such as bubble or crack) inside the solder paste usually has a lower thermal conductivity , the temperature rise rate of these areas is different from that of the normal area, and these temperature differences are captured by the infrared thermal imager, and the defect distribution of the solder paste surface is judged by analyzing the thermal image data . The processed ultrasonic signal and the transient temperature distribution The spatial and temporal registration process is performed to ensure that they are in the same coordinate system for data fusion. The SIFT (Scale-Invariant Feature Transform) based image matching algorithm is used to align the thermal image data and ultrasonic signal data in space, and the geometric distortion of the two is corrected by affine transformation, so that the thermal image data and ultrasonic signals match in the same physical area. The time series of ultrasonic signals and thermal image data are aligned by a time synchronization algorithm, i.e. at each laser pulse trigger, the time stamp of the ultrasonic signal and the time stamp of the thermal image data are recorded, and the non-synchronized data is time-matched by an interpolation method to ensure that the data points of the two correspond to the same laser irradiation time on the time axis. The registered data set is composed of ultrasonic signals and transient temperature distribution , which not only contains the ultrasonic response information of the thin-wall layer of the solder paste, but also combines the thermal physical properties of the material to form a complete thermal-acoustic coupling response data set.

[0020] In one example, the thin-wall layer of the solder paste with quantum dots mixed in advance is scanned by fluorescence lifetime imaging, and a fluorescence parameter data set is obtained, including: CdSe / ZnS core-shell structure quantum dots mixed in advance in the thin-wall layer of the solder paste are excited and irradiated by a picosecond pulsed laser to obtain a fluorescence signal generated by the quantum dots; The fluorescence signal is separated by a dichroic filter and received and converted by a photomultiplier tube to obtain a fluorescence electrical signal; The fluorescence electrical signal is time-series collected by a time correlation single photon counting technology to obtain a fluorescence decay curve; The fluorescence decay curve is least squares fitted by a three-exponential model to obtain the fluorescence lifetime parameters of the quantum dots in different environments; The fluorescence lifetime parameters, fluorescence intensity and emission peak position information are spatially filtered and denoised to obtain a fluorescence parameter data set.

[0021] In this example, a high-energy-density picosecond pulsed laser is used for excitation and irradiation, the central wavelength of the laser is set to nm, the pulse width is ps, and the repetition frequency is MHz, to ensure that the laser can provide enough energy to excite the quantum dots inside the solder paste and generate a measurable fluorescence signal. When the laser irradiates the thin-wall layer of the solder paste, the CdSe / ZnS core-shell structure quantum dots mixed in advance absorb photons and transition to the excited state, and after a certain lifetime, they release a fluorescence signal through radiative transition. The emission wavelength range of the fluorescence is nm, and the fluorescence lifetime is Affected by the microenvironment of solder paste, the defective area of ​​solder paste is identified by analyzing the fluorescence lifetime. In order to accurately detect the fluorescence signal generated by quantum dots, a dichroic filter is used to separate the fluorescence signal by wavelength. The cutoff wavelength of the filter is set to The fluorescence signal after passing through the filter is received and converted by a high-sensitivity photomultiplier tube (PMT). The photoelectric conversion efficiency of the PMT is The wavelength is greater than 40%, which can ensure efficient detection of weak fluorescence signals and convert the optical signal into a fluorescent electrical signal through an internal high-gain amplification circuit. The electrical signal contains the timing information of the fluorescence intensity, and its amplitude is affected by the fluorescence quantum yield, the local microenvironment and the concentration of quantum dots, so it needs further processing to extract key parameters. After the fluorescence electrical signal is acquired, the signal is collected in a time-series manner with high time resolution by using time-correlated single-photon counting technology. This method is based on the statistical distribution of photon arrival time and can accurately record the fluorescence decay process. Since the fluorescence lifetime is affected by the local environment, when the quantum dots are in a normal solder paste matrix, the main lifetime component The fluorescence lifetime is about 15-20 ns. When the quantum dots are in defective areas (such as bubbles, cracks or oxidation areas), changes in the local microenvironment will cause the fluorescence lifetime to be shortened to 5-10 ns, accompanied by a blue shift (5-15 nm) or red shift (10-20 nm) of the fluorescence emission peak. By analyzing the fluorescence lifetime distribution, the defective areas of the solder paste thin wall layer can be effectively identified. In order to obtain the fluorescence lifetime parameters, the fluorescence decay curve is analyzed. Perform least squares fitting calculation to optimize the fluorescence lifetime parameters and the corresponding amplitude coefficient The fitting process uses a nonlinear least squares optimization algorithm and is based on the goodness of fit. As a guideline to ensure the accuracy of the calculation results. After the calculation is completed, the fluorescence lifetime spatial distribution map is constructed. ,in is the two-dimensional coordinate of the solder paste thin wall layer. This distribution diagram can intuitively show the difference in fluorescence lifetime in different areas of the solder paste, thus providing basic data for defect identification. In order to improve the signal-to-noise ratio of fluorescence data, spatial filtering and denoising are performed on the fluorescence lifetime parameters, fluorescence intensity and emission peak position information, and Gaussian filtering is performed on the fluorescence intensity data to reduce the impact of random noise on the detection results. and fluorescence emission peak position information An adaptive mean filter method is used for denoising. This method dynamically adjusts the filter weight based on the signal distribution characteristics of the local area to maximize the preservation of the characteristic information of the defect area. After data preprocessing and feature extraction, a fluorescence parameter dataset containing fluorescence lifetime parameters, fluorescence intensity distribution, and emission peak position information is obtained.

[0022] In one example, a polarization spectroscopy dataset, a thermal-acoustic coupling response dataset, and a fluorescence parameter dataset are fed into an attention-guided deformable convolutional network for feature extraction and fusion, outputting a comprehensive detection report including: The polarization spectrum matrix in the polarization spectrum dataset is converted into a two-dimensional image matrix through wavelength channel combination and polarization state mapping processing; The acoustic impedance distribution and transient temperature distribution map in the thermal-acoustic coupled response dataset are projected into a two-dimensional feature map; Reconstruct the fluorescence lifetime, intensity and emission peak position information in the fluorescence parameter data set into a two-dimensional feature distribution map; Based on the affine transformation algorithm, the spatial coordinates of the two-dimensional image matrix, the two-dimensional feature map and the two-dimensional feature distribution map are aligned to obtain a multimodal image dataset corresponding to the same physical area; The multimodal image dataset is input into the backbone feature extraction network of the attention-guided deformable convolutional network for depthwise separable convolution processing to obtain the initial feature map containing the thin-walled solder paste structure information; The initial feature map is input into the polarization feature branch, acoustic feature branch, and fluorescence feature branch for parallel feature extraction. This results in three sets of channel feature maps representing different physical properties of thin-wall solder paste. The polarization feature branch extracts polarization degree and dual-wavelength difference information, the acoustic feature branch extracts sound velocity and spectrum characteristics, and the fluorescence feature branch extracts fluorescence lifetime and spectral shift information. The weights of the three groups of channel feature maps are dynamically adjusted through the channel attention mechanism to obtain weighted feature maps. The receptive field shape of the weighted feature maps is automatically adjusted through the deformable convolution module to obtain feature maps that adapt to thin-wall boundaries. The feature map adapted to the thin-wall boundary is subjected to dimensionality reduction processing to obtain a comprehensive feature map containing comprehensive feature information of the thin-wall layer solder paste. The comprehensive feature map is the feature fusion result; Based on the comprehensive feature map, multi-threshold self-adjusting segmentation and depth estimation calculation are performed on the thin-walled solder paste to obtain a defect mask that characterizes the defect location of the thin-walled solder paste and a three-dimensional defect model that characterizes the defect depth distribution; The defect mask and 3D defect model are combined with the solder paste metallurgical feature library for multi-task learning analysis to obtain a comprehensive inspection report.

[0023] In this example, for the polarization spectrum dataset, the polarization spectrum matrix Contains multiple wavelengths and polarization angle of each pixel representing a certain spatial position on the sample surface. Due to the high-dimensional characteristics of the original data with multiple wavelengths and polarization angles, it is converted into a two-dimensional image matrix through wavelength channel combination and polarization state mapping. Through wavelength channel combination, the polarization degree data of different wavelengths are merged. The merging method includes principal component analysis dimension reduction to extract the most representative wavelength channel, or using a feature selection method based on mutual information to retain the most important wavelength information for thin-wall layer solder paste detection, thereby reducing data redundancy and improving computational efficiency. In order to optimize the polarization state mapping, the polarization information of each pixel is vectorized and converted, and the polarization ellipse parameters are calculated, including the polarization degree and polarization angle , and then these data are mapped into pseudo-color images, in which can be used to distinguish the optical anisotropy of different materials, and is used to detect the orientation and stress distribution of the thin-wall layer solder paste. This conversion process obtains a two-dimensional polarization spectral feature map representing the polarization characteristics of the solder paste surface. For the thermo-acoustic coupling response dataset, the core information comes from the acoustic impedance distribution and transient temperature distribution map . Among them, the acoustic impedance data is essentially a three-dimensional data matrix representing the acoustic characteristics at different depth layers , so depth projection is performed to convert it into a two-dimensional feature map. In this process, the maximum projection method is used, that is, for each position, the layer with the maximum acoustic impedance is selected, so as to highlight the high acoustic impedance features of the defect area. Through weighted average projection, the acoustic impedance information of each pixel point is the weighted sum of multiple depth layers, improving the stability of the data. For the transient temperature distribution , the time dimension represents the temperature evolution process after laser heating, so it is converted through time integration or peak extraction. For example, the maximum temperature rise projection is adopted, that is, the maximum value in the direction is taken , to obtain the strongest response area of the solder paste surface temperature after laser action. The final two-dimensional feature map includes and , representing the acoustic and thermal characteristics of the solder paste, respectively. For the fluorescence parameter dataset, the core variables include fluorescence lifetime , fluorescence intensity and emission peak position Fluorescence lifetime is measured based on time-resolved single-photon counting method, while fluorescence intensity relies on the integral value of fluorescence signal. In order to reconstruct the two-dimensional feature distribution map, spatial filtering is performed on the fluorescence lifetime data to reduce noise and improve detection accuracy. For example, Gaussian filtering is used to smooth the fluorescence lifetime distribution to eliminate random errors of individual pixel points. At the same time, for the fluorescence intensity data , normalization processing is performed to keep the fluorescence signals of all pixel points within the same dynamic range. For the fluorescence emission peak position Because its value is significantly affected by environmental changes, it is denoised using mean filtering. A fluorescence peak shift image is generated, which provides information about the spectral changes in the defect area. These fluorescence feature data are combined into a two-dimensional feature distribution map containing fluorescence lifetime, intensity, and spectral shift information. This image is used to analyze the microchemical environment and material properties of the solder paste thin layer. The polarization spectrum feature map, the thermoacoustic coupling feature map, and the fluorescence feature map are spatially aligned to align them to the same physical region. An affine transformation algorithm is used to accurately register them. Feature point matching methods (such as SIFT or ORB) are used to find corresponding key points in different images. The affine transformation matrix is ​​then calculated using a least squares method or RANSAC (random sample consensus) optimization method. This affine transformation is used to map all datasets to a unified spatial coordinate system, resulting in a registered multimodal image dataset. This multimodal image dataset is then fed into an attention-guided deformable convolutional network for depthwise separable convolution. The backbone feature extraction network of this network uses depthwise separable convolutional layers to extract high-level features from the input data while reducing computational complexity. Depthwise separable convolution not only improves convolution efficiency but also effectively preserves the spatial characteristics of the data. This network extracts an initial feature map containing structural information about the solder paste, encompassing both its microstructure and physical properties. After feature extraction, this initial feature map is fed into three parallel feature branches: a polarization feature branch, an acoustic feature branch, and a fluorescence feature branch. Each branch extracts features specific to a specific physical property. For example, the polarization feature branch extracts degree of polarization and dual-wavelength difference information, revealing surface anisotropy of the solder paste; the acoustic feature branch extracts sound velocity and spectral characteristics, aiding in analyzing acoustic propagation characteristics and material uniformity within the solder paste; and the fluorescence feature branch extracts fluorescence lifetime and spectral shift information, providing detailed information on content distribution and defects within the solder paste. Each feature branch extracts a different channel feature map through convolutional layers, resulting in three channel feature maps. The weights of these three channel feature maps are dynamically adjusted using a channel attention mechanism. By weighting each channel, the network automatically learns the importance of different physical properties in solder paste defect detection. This attention mechanism strengthens focus on key features and improves the network's recognition capabilities. A deformable convolution module automatically adjusts the receptive field shape of the weighted feature map to accommodate the irregular boundaries of the thin solder paste layer. By introducing an additional offset field, deformable convolution flexibly adjusts the receptive field shape of the convolution kernel, thereby improving the network's adaptability to complex boundaries. After deformable convolution, the resulting feature map undergoes dimensionality reduction. By using a 1×1 convolution operation to reduce the number of channels in the feature map and using pooling to reduce the spatial dimension, a comprehensive feature map containing comprehensive feature information of the thin solder paste layer is extracted. This comprehensive feature map is the final result of feature fusion, integrating information from all different physical properties.Based on the comprehensive feature map, multi-threshold self-adjusting segmentation and depth estimation calculation are performed to obtain a defect mask and a three-dimensional defect model. In the segmentation process, an adaptive threshold algorithm is used to accurately divide the defect area. By combining the U-Net architecture in the deep learning network and the region growing method, high-precision defect area extraction is achieved. Based on three-dimensional acoustic impedance data and fluorescence information, the system performs depth estimation to calculate the depth distribution of the solder paste defects and generates a three-dimensional defect model. After completing the defect segmentation and depth estimation, the defect mask and three-dimensional defect model are analyzed with the solder paste metallographic feature library through multi-task learning to obtain a comprehensive detection report. The report contains the type, severity, size, shape, and depth information of the defects, and presents the defect distribution map and three-dimensional structure model through a visualization tool.

[0024] In one example, based on the comprehensive feature map, multi-threshold self-adjusting segmentation and depth estimation calculation are performed on the thin-wall layer solder paste to obtain a defect mask representing the location of the thin-wall layer solder paste defects and a three-dimensional defect model representing the depth distribution of the defects, including: The comprehensive feature map is input into a multi-scale feature pyramid for processing to obtain layered feature maps that can capture thin-wall layer solder paste defects of different sizes; The layered feature maps are input into a U-Net segmentation network for feature extraction and upsampling processing to obtain semantic feature representations of the thin-wall layer solder paste defects; A boundary perception module is introduced into the U-Net segmentation network to calculate the feature gradient amplitude and direction of the thin-wall area based on the semantic feature representations, obtaining three sets of channel output maps, including a defect probability map, a boundary map, and a direction map; An adaptive threshold function is used to dynamically calculate the threshold of the defect probability map to obtain a seed point set for region growing; The seed point set is expanded through region growing and level set methods, and is constrained by the boundary map to obtain a defect mask representing the location of the thin-wall layer solder paste defects; The defect mask and the acoustic impedance distribution in the thermo-acoustic coupling response dataset are input into a depth estimation network for three-dimensional reconstruction calculation to obtain a three-dimensional defect model representing the depth distribution of the defects.

[0025] In this example, the comprehensive feature map is input into a multi-scale feature pyramid for processing to obtain layered feature maps that can capture thin-wall layer solder paste defects of different sizes. In the multi-scale feature pyramid, the input comprehensive feature map is divided into multiple scale levels, each scale corresponding to a different resolution layer, where high-resolution layers are used to capture detailed information of small defects, and low-resolution layers are used to extract global structural features. For each scale , a Gaussian pyramid downsampling function Down-sampling is performed so that each hierarchical feature map has different spatial resolutions, while introducing Laplacian operator in the down-sampling process to enhance edge and texture features, thus ensuring the integrity of features at different scales. Through this step, the multi-scale feature pyramid effectively integrates information at different spatial scales, enabling the system to simultaneously recognize large-scale defects and tiny solder paste unevenness. The hierarchical feature maps output by the multi-scale feature pyramid are input into the U-Net segmentation network for feature extraction and up-sampling processing to obtain semantic feature representations of the solder paste thin-wall layer defects. In the encoding stage of the U-Net network, multiple deep residual blocks are used, each containing two convolutional layers and a cross-layer skip connection to ensure that the network can learn deep features while maintaining the integrity of the information. In the decoding stage, deconvolution operation is used for step-by-step up-sampling, and the features from the encoding stage are mapped to high-resolution layers through skip connections to recover the spatial information of the defect area. To improve the network's ability to recognize the boundaries of the solder paste thin-wall layer, a boundary perception module is introduced into the U-Net network. This module calculates the gradient amplitude and direction information to enhance the boundary perception ability of the thin-wall area. By calculating the gradient information of each pixel, three sets of channel output maps are constructed, including a defect probability map , a boundary map , and a direction map . The defect probability map is normalized to [0, 1] by the Sigmoid function, representing the probability of each pixel belonging to a defect; the boundary map is used to locate the boundaries of the defect area and is post-processed using Canny edge detection to optimize the edge information; the direction map is used to represent the main direction information of the defect, and the main extension direction of the defect is determined by calculating the gradient direction angle distribution. To optimize the recognition results of the defect area, an adaptive threshold function is used to dynamically calculate the threshold of the defect probability map to obtain a set of seed points for region growing. This threshold calculation method dynamically adjusts the threshold parameter based on local statistical characteristics, so that the threshold changes adaptively in different regions, ensuring that the seed point selection for region growing has better robustness. After obtaining the seed point set, region growing and level set methods are used for region expansion, and the boundary map is combined for constraint to accurately segment the defect area. This method effectively expands the initial seed region and ultimately forms a complete defect area. To ensure the accuracy of the region boundary, a level set method is used for boundary optimization, which optimizes the defect boundary through an adaptive evolution method and finally generates a defect mask . After extracting the defect mask, the acoustic impedance distribution Defect depth estimation is performed. The defect mask is spatially registered with the acoustic impedance data to ensure consistency of coordinate systems. Based on the ultrasonic propagation time and impedance characteristics, a depth estimation model is constructed. The model infers the depth distribution of the defect region by calculating the time delay of the ultrasonic wave based on the propagation characteristics of the ultrasonic wave, and finally generates a three-dimensional defect model In one example, the defect mask and the three-dimensional defect model are analyzed by multi-task learning with the tin paste metallurgical feature library to obtain a comprehensive detection report, including: Geometric descriptors and texture descriptors are extracted from the defect mask, depth distribution features are extracted from the three-dimensional defect model, acoustic descriptors are extracted from the thermal-acoustic coupling response dataset, and fluorescence descriptors are extracted from the fluorescence parameter dataset to obtain a complete set of multi-dimensional descriptor set that characterizes the defect characteristics of the thin-wall layer tin paste; The multi-dimensional descriptor set is input into a 5-layer fully connected network for feature transformation to obtain a target feature vector; The target feature vector is input into a defect type classification branch, a defect severity evaluation branch, and a defect depth positioning branch for multi-task parallel calculation to obtain an initial defect classification result; The initial defect classification result is compared with the tin paste metallurgical feature library for similarity calculation to obtain a matching result, and the matching result is integrated through an integrated decision strategy to obtain defect type and severity information; Defect quantitative evaluation data including size dispersion, depth distribution, and morphological complexity are calculated based on the defect mask and the three-dimensional defect model, and the defect quantitative evaluation data are graded based on a pre-set threshold to obtain a defect severity grading result; The defect type, severity information, defect quantitative evaluation data, and defect severity grading result are integrated to generate a grading warning information, forming a comprehensive detection report including defect distribution map, three-dimensional structure visualization, and evaluation data.

[0026] In this example, based on the defect mask Geometric features are extracted. Geometric descriptors include defect area , perimeter , aspect ratio , and shape factor , which measure the morphological complexity of the defect. If is close to 1, it indicates that the defect shape is close to a circle, and if it is much smaller than 1, it indicates that the defect has a more complex boundary. At the same time, texture descriptors are calculated from the gray level co-occurrence matrix of the defect region, such as contrast , entropy , and uniformity , which are used to characterize the surface texture features of the defect region. Based on the geometric and texture features, depth distribution features are extracted from the three-dimensional defect model The key parameters of defect depth distribution include the maximum depth , average depth and depth change rate Higher The value indicates that the defect surface is relatively undulating. Acoustic descriptors are extracted from the thermal-acoustic coupled response data set, including the mean acoustic impedance , ultrasonic attenuation coefficient and the main resonant frequency , which is used to measure the material impedance change in the defect area, while the ultrasonic attenuation coefficient is measured by the signal amplitude At a specific frequency The fluorescence descriptors, including the mean fluorescence lifetime, are extracted from the fluorescence parameter dataset. , life distribution standard deviation and spectral shift , which can reflect the degree of change in the fluorescence lifetime of the defect area. If the value is large, it indicates that there is a significant change in the material environment of the defect area, such as oxidation or doping inhomogeneity. After constructing a complete set of multidimensional descriptors, it is input into a 5-layer fully connected network for feature transformation to obtain the target feature vector. In this neural network, the output of each layer of neurons is given by a nonlinear transformation, where is the weight matrix, is the bias term, Represents the activation function. The network transforms the input descriptor set layer by layer to learn a more compact and effective feature representation and finally generates the target feature vector . The target feature vector is input into the defect type classification branch, defect severity assessment branch and defect depth location branch for multi-task parallel calculation to obtain the initial defect classification result. The defect type classification branch uses a Softmax classifier to calculate the probability of each defect type, the defect severity assessment branch predicts the severity of the defect through a regression model, and the defect depth location branch calculates the depth coordinate of the defect center based on the depth estimation network. The initial defect classification result is calculated for similarity with the solder paste metallurgical feature library to obtain a matching result. The matching results are integrated through an integrated decision strategy to obtain the optimal defect type and severity information. Based on the defect mask and the three-dimensional defect model, the defect quantitative evaluation data including size discreteness, depth distribution and morphological complexity are calculated, and these data are graded based on the preset threshold. Based on these indicators, the severity of the defects is divided according to the preset threshold, such as mild (S level), medium (M level), severe (L level) and extreme (X level) to form a complete defect severity grading result. Integrate defect type, severity information, defect quantitative assessment data and defect severity grading results to generate graded warning information and form a comprehensive inspection report including defect distribution map, 3D structure visualization and assessment data.

[0027] The embodiment also includes thin-wall layer solder paste defect timing evolution prediction and dynamic monitoring based on the comprehensive detection report, specifically including: repeatedly performing the detection steps on the thin-wall layer solder paste of the same PCB at different time points to obtain a defect data sequence containing a time dimension, each time point forming a set of defect feature vectors; time domain reconstruction is performed on the defect data sequence, the defect mask and the three-dimensional defect model at different time points are registered and superimposed through the sliding window method to obtain a four-dimensional space-time model representing the dynamic evolution process of the defect; the rate of change of the defect geometric parameters, the morphology evolution speed and the time-varying characteristics of the thermo-acoustic-optical characteristics are extracted based on the four-dimensional space-time model to construct a defect dynamic feature set containing 10 time sequence descriptors; the defect dynamic feature set is input into a long short-term memory network for time sequence learning to train a defect development trend prediction model, the prediction model includes 3 layers of bidirectional LSTM units and 2 layers of time attention mechanisms; based on the defect development trend prediction model, the future state of the defect is predicted at T+1 to T+5 time points in multiple steps to generate prediction data containing the change trend of the defect size, depth and severity; by comparing the defect parameter change rate in the prediction data with the preset critical threshold, a risk growth curve is constructed, and the predicted time for the defect to reach the danger level is calculated; according to the relationship between the predicted time and the production process node, the warning level is dynamically adjusted and early warning signals containing three time dimensions are generated: immediate warning, short-term warning and long-term warning; based on the defect dynamic feature set and the prediction data, an association model of process parameters and defect formation is established, and a genetic algorithm is used to optimize and solve the optimal process parameter combination to generate a solder paste printing process adaptive adjustment suggestion; the early warning signal and the process adaptive adjustment suggestion are transmitted to the production control system in real time to form a closed-loop feedback control mechanism; a thin-wall layer solder paste defect knowledge graph is constructed, the correlation between the newly detected defect features and the historical data is analyzed, the potential rules of defect formation are mined through a graph attention network, and the prediction accuracy of the defect development trend prediction model is continuously updated and optimized.

[0028] Referring to Figure 2 The embodiment provides a thin-wall layer defect detection device for solder paste coating, which comprises: A scanning module 1 is used for performing double-wavelength polarized light scanning on the thin-wall layer solder paste on a PCB to obtain a polarized spectrum data set; An acquisition module 2 is used for performing pulsed laser irradiation and ultrasonic signal acquisition on the thin-wall layer solder paste to obtain a thermo-acoustic coupling response data set; An imaging module 3 is used for performing fluorescence lifetime imaging scanning on the thin-wall layer solder paste mixed with quantum dots in advance to obtain a fluorescence parameter data set; An output module 4 is used for inputting the polarized spectrum data set, the thermo-acoustic coupling response data set and the fluorescence parameter data set into an attention-guided deformable convolution network for feature extraction and fusion, and outputting a comprehensive detection report.

[0029] In this embodiment, for specific implementation of each unit in the above device embodiment, please refer to the description in the above method embodiment, which will not be repeated here.

[0030] It should be noted that in this paper, the term "includes", "contains" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, device, article or method including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or includes elements inherent to such process, device, article or method. Without more limitations, the element defined by the statement "includes a" does not exclude the existence of other identical elements in the process, device, article or method including the element.

[0031] The above is only the preferred embodiment of the present application, and does not limit the patent scope of the present application, and any equivalent structure or equivalent process transformation using the content of the present application specification and drawings, or direct or indirect application in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A method for detecting defects in thin-walled layers coated with solder paste, characterized in that: include: Perform dual-wavelength polarized light scanning on the thin layer of solder paste on the PCB to obtain a polarization spectrum data set; performing pulsed laser irradiation and ultrasonic signal acquisition on the thin-walled solder paste layer to obtain a thermal-acoustic coupling response data set; Performing fluorescence lifetime imaging scanning on the thin-walled solder paste pre-mixed with quantum dots to obtain a fluorescence parameter data set; The polarization spectrum dataset, the thermal-acoustic coupling response dataset, and the fluorescence parameter dataset are input into an attention-guided deformable convolutional network for feature extraction and fusion, and a comprehensive detection report is output.

2. The method for detecting defects in thin-walled layers of solder paste coating according to claim 1, wherein: The method of performing dual-wavelength polarized light scanning on the thin-walled solder paste layer on the PCB to obtain a polarization spectrum data set includes: The lasers generated by the blue laser and the near-infrared laser are combined by a dichroic beam splitter to obtain a dual-wavelength beam, and the dual-wavelength beams are polarization-separated by a polarization beam splitter to obtain orthogonal polarized light; The orthogonal polarized light is converted into a circularly polarized light by a quarter wave plate, and the circularly polarized light is focused by a collimating and focusing system to obtain a surface light spot focused on the thin-walled layer of solder paste; Performing phase detection on the light signal reflected from the surface of the thin-walled solder paste layer using a polarization analyzer array to obtain polarization state change information of the dual-wavelength channel; Dark background subtraction, non-uniformity correction and polarization degree calculation are performed on the polarization state change information of the dual-wavelength channel to obtain a polarization spectrum data set.

3. The method for detecting defects in thin-walled layers coated with solder paste according to claim 1, wherein: The thin-walled solder paste layer is subjected to pulsed laser irradiation and ultrasonic signal acquisition to obtain a thermal-acoustic coupling response data set, including: Scanning and irradiating the thin-walled solder paste layer with a laser beam generated by a Q-switched pulsed laser to obtain ultrasonic waves generated by a thermoelastic effect, and receiving and collecting the ultrasonic waves generated on the surface of the thin-walled solder paste layer through a piezoelectric sensor array to obtain an original ultrasonic signal; The original ultrasonic signal is subjected to signal enhancement and filtering processing by a low-noise preamplifier and a bandpass filter to obtain a processed ultrasonic signal, and the surface of the thin-walled solder paste layer is subjected to synchronous thermal imaging acquisition by an infrared thermal imager to obtain a transient temperature distribution map; The processed ultrasonic signal and the transient temperature distribution map are registered to obtain a thermal-acoustic coupling response data set.

4. The method for detecting defects in thin-walled layers of solder paste coating according to claim 1, wherein: The thin-walled solder paste pre-mixed with quantum dots is subjected to fluorescence lifetime imaging scanning to obtain a fluorescence parameter data set, including: Using a picosecond pulse laser to excite and irradiate the CdSe / ZnS core-shell structure quantum dots pre-mixed in the thin-walled solder paste to obtain a fluorescence signal generated by the quantum dots; The fluorescence signal is wavelength-separated by a dichroic filter and received and converted by a photomultiplier tube to obtain a fluorescence electrical signal; The fluorescence electrical signal is collected in time series by using a time-correlated single photon counting technique to obtain a fluorescence decay curve; The fluorescence decay curve is fitted by a three-exponential model using the least squares method to obtain the fluorescence lifetime parameters of the quantum dots under different environments; The fluorescence lifetime parameter, fluorescence intensity and emission peak position information are spatially filtered and denoised to obtain a fluorescence parameter data set.

5. The method for detecting defects in thin-walled layers of solder paste coating according to claim 1, wherein: The polarization spectrum dataset, the thermal-acoustic coupling response dataset, and the fluorescence parameter dataset are input into an attention-guided deformable convolutional network for feature extraction and fusion, and a comprehensive detection report is output, including: Converting the polarization spectrum matrix in the polarization spectrum dataset into a two-dimensional image matrix through wavelength channel combination and polarization state mapping processing; Projecting the acoustic impedance distribution and the transient temperature distribution map in the thermal-acoustic coupling response data set into a two-dimensional feature map; Reconstructing the fluorescence lifetime, intensity and emission peak position information in the fluorescence parameter data set into a two-dimensional feature distribution map; Performing spatial coordinate registration on the two-dimensional image matrix, the two-dimensional feature map, and the two-dimensional feature distribution map based on an affine transformation algorithm to obtain a multimodal image dataset corresponding to the same physical area; Inputting the multimodal image dataset into the backbone feature extraction network of the attention-guided deformable convolutional network for depthwise separable convolution processing to obtain an initial feature map containing thin-walled solder paste structure information; The initial feature map is input into the polarization feature branch, the acoustic feature branch, and the fluorescence feature branch for parallel feature extraction, thereby obtaining three sets of channel feature maps that characterize different physical properties of the thin-walled solder paste. The polarization feature branch extracts polarization degree and dual-wavelength difference information, the acoustic feature branch extracts sound velocity and spectrum characteristics, and the fluorescence feature branch extracts fluorescence lifetime and spectral shift information. Dynamically adjust the weights of the three groups of channel feature maps through the channel attention mechanism to obtain a weighted feature map, and automatically adjust the receptive field shape of the weighted feature map through the deformation convolution module to obtain a feature map adapted to the thin wall boundary; Performing dimensionality reduction processing on the feature map adapted to the thin-wall boundary to obtain a comprehensive feature map containing comprehensive feature information of the thin-wall layer solder paste, wherein the comprehensive feature map is the feature fusion result; Based on the comprehensive feature map, multi-threshold self-adjusting segmentation and depth estimation calculation are performed on the thin-walled solder paste layer to obtain a defect mask representing the defect position of the thin-walled solder paste layer and a three-dimensional defect model representing the defect depth distribution; The defect mask and the three-dimensional defect model are subjected to multi-task learning analysis with a solder paste metallurgical feature library to obtain a comprehensive inspection report.

6. The method for detecting defects in thin-walled layers coated with solder paste according to claim 5, wherein: The method of performing multi-threshold self-adjusting segmentation and depth estimation calculation on the thin-walled solder paste layer based on the comprehensive feature map to obtain a defect mask characterizing the defect position of the thin-walled solder paste layer and a three-dimensional defect model characterizing the defect depth distribution includes: Inputting the comprehensive feature map into a multi-scale feature pyramid for processing to obtain a layered feature map capable of capturing thin-walled solder paste defects of different sizes; Inputting the hierarchical feature map into a U-Net segmentation network for feature extraction and upsampling processing to obtain a semantic feature representation of the solder paste thin wall layer defect; Introducing a boundary perception module into the U-Net segmentation network, calculating the feature gradient amplitude and direction of the thin-walled area according to the semantic feature representation, and obtaining three sets of channel output maps, the three sets of channel output maps including a defect probability map, a boundary map, and a direction map; Performing dynamic threshold calculation on the defect probability map using an adaptive threshold function to obtain a set of seed points for region growth; Expanding the seed point set by using region growing and level set methods, and constraining it in combination with the boundary map, to obtain a defect mask that characterizes the location of defects in the thin-walled solder paste; The defect mask and the acoustic impedance distribution in the thermal-acoustic coupling response data set are input into a depth estimation network for three-dimensional reconstruction calculation to obtain a three-dimensional defect model that characterizes the defect depth distribution.

7. The method for detecting defects in thin-walled layers of solder paste coating according to claim 5, characterized in that: The defect mask and the three-dimensional defect model are subjected to multi-task learning analysis with the solder paste metallurgy feature library to obtain a comprehensive inspection report, including: Extracting geometric descriptors and texture descriptors from the defect mask, extracting depth distribution features from the three-dimensional defect model, extracting acoustic descriptors in combination with the thermal-acoustic coupling response dataset, and extracting fluorescence descriptors in combination with the fluorescence parameter dataset, thereby obtaining a multidimensional descriptor set that fully characterizes the defects of thin-walled solder paste; Inputting the multidimensional descriptor set into a 5-layer fully connected network for feature transformation to obtain a target feature vector; Inputting the target feature vector into the defect type classification branch, the defect severity assessment branch, and the defect depth location branch for multi-task parallel calculation respectively to obtain an initial defect classification result; Performing similarity calculation on the initial defect classification result and the solder paste metallurgical feature library to obtain a matching result, and integrating the matching result through an integrated decision strategy to obtain defect type and severity information; Calculating defect quantitative assessment data including size dispersion, depth distribution, and morphological complexity based on the defect mask and the three-dimensional defect model, and grading the defect quantitative assessment data based on a preset threshold to obtain a defect severity grading result; The defect type, the severity information, the defect quantitative assessment data and the defect severity grading result are integrated to generate graded warning information, forming a comprehensive inspection report including a defect distribution map, three-dimensional structure visualization and assessment data.

8. A device for detecting defects in thin-walled layers coated with solder paste, characterized in that: For implementing the steps of the method according to any one of claims 1 to 7, the device for detecting defects in thin-walled layers coated with solder paste comprises: The scanning module is used to perform dual-wavelength polarized light scanning on the thin-walled solder paste layer on the PCB to obtain a polarization spectrum data set; an acquisition module, configured to perform pulsed laser irradiation and ultrasonic signal acquisition on the thin-walled solder paste layer to obtain a thermal-acoustic coupling response data set; An imaging module is used to perform fluorescence lifetime imaging scanning on the thin-walled solder paste pre-mixed with quantum dots to obtain a fluorescence parameter data set; The output module is used to input the polarization spectrum dataset, the thermal-acoustic coupling response dataset and the fluorescence parameter dataset into the attention-guided deformable convolutional network for feature extraction and fusion, and output a comprehensive detection report.

Images