Method for detecting welding quality of flexible circuit board of audio equipment

Abstract

The invention discloses a method for detecting the welding quality of a flexible circuit board of audio equipment, which relates to the technical field of electronic manufacturing quality detection, and comprises the following steps: forming a standardized welding spot image sequence and a temperature change curve by synchronously acquiring a multispectral image and infrared thermal imaging time sequence data of a welding area and performing registration and enhancement processing. And geometric texture features of the image and heat conduction and cooling features of the temperature curve are extracted, a multi-dimensional feature descriptor is constructed through fusion, and a deep learning network is input to obtain the welding spot quality grade and defect probability. And in combination with the defect probability and the temperature curve, positioning the defect and analyzing the process stage generated by the defect, and finally generating an evaluation report containing repair suggestions. According to the method, multi-modal dynamic detection is realized, and the identification accuracy of hidden defects and the traceability of process problems are effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of electronic manufacturing quality inspection technology, specifically a method for inspecting the welding quality of flexible circuit boards for audio equipment. Background Technology

[0002] Currently, the soldering quality inspection of flexible circuit boards in audio equipment mainly relies on automated optical inspection or manual visual inspection. These methods are based on two-dimensional images under visible light, judging the soldering quality by comparing preset static features such as solder joint contours, positions, and gloss. Some high-end inspections introduce X-ray imaging to observe structural defects such as holes and cracks inside the solder joints. However, these technologies all perform static analysis on the final state after soldering, and the data sources and information dimensions they rely on are relatively limited.

[0003] Existing technical solutions have shortcomings. Static optical or X-ray inspection cannot capture the dynamic physical changes during the welding process. This results in inspection remaining at the superficial level, with low identification rates for latent defects closely related to the thermal process, such as cold solder joints and poor welds, and it is completely impossible to determine whether the defect was generated during the heating, melting, or cooling stage. When faced with tiny, dense, and material-specific solder joints on flexible circuit boards, the limited information from a single image modality is insufficient to support the accurate classification and root cause tracing of complex defects.

[0004] There is a need for a detection method that can integrate dynamic information from the welding process with multi-dimensional static characteristics of weld joints. It is also necessary to address the technical challenges of evolving from single static detection to multimodal dynamic analysis, in order to achieve a more in-depth assessment of welding quality and more instructive fault diagnosis for production processes. Summary of the Invention

[0005] This invention aims to solve at least one of the technical problems existing in the prior art; Therefore, this invention proposes a method for inspecting the welding quality of flexible circuit boards in audio equipment, comprising: Acquire multispectral scanning image data and infrared thermal imaging time-series data of the flexible circuit board surface of the audio device under test to form a set of original welding area multimodal data; Image enhancement and data registration are performed on the original welding area multimodal data set to generate a standardized weld point feature image sequence and corresponding temperature distribution time-series curve; Extract the geometric morphology feature vector and texture distribution feature vector of the solder joint from the standardized solder joint feature image sequence, and extract the heat conduction feature vector and cooling feature vector of the solder joint from the temperature distribution time series curve; The geometric shape feature vector, texture distribution feature vector, heat conduction feature vector and cooling feature vector are fused to construct a multi-dimensional solder joint quality feature descriptor. The multidimensional weld joint quality feature descriptor is input into a preset deep learning classification network, and the deep learning classification network outputs the quality level label and defect type probability distribution for each weld joint. Based on the probability distribution of the defect type and the time series curve of the temperature distribution, the specific location of the welding defect is located and the time stage of defect occurrence is analyzed. Based on the quality grade label and defect location information of each weld point, a welding quality assessment report containing repair priorities and repair recommendations is generated.

[0006] Furthermore, the image enhancement and data registration processing of the original welding area multimodal data set to generate a standardized weld point feature image sequence and corresponding temperature distribution time-series curve includes: Denoising processing is performed on the multispectral scan image data, and an adaptive filtering algorithm is used to eliminate image noise and enhance the contrast between the solder joints and the background. Temperature calibration is performed on the time-series infrared thermal imaging data to convert the original thermal imaging grayscale values ​​into precise temperature values ​​and establish a time-temperature correspondence. Identify the contour location of solder joints in multispectral images and establish a local coordinate system with the center of the solder joint as the reference. Infrared thermal imaging data is mapped to the local coordinate system to ensure that the position of each solder joint corresponds precisely in the thermal imaging data and multispectral image; The registered multispectral image sequence is subjected to size normalization processing to adjust all solder joint images to a uniform resolution, forming the standardized solder joint feature image sequence. Temperature change data for each solder joint is extracted from the registered thermal imaging data and arranged in chronological order to form the corresponding temperature distribution time-series curve.

[0007] Further, the extraction of geometric morphological feature vectors and texture distribution feature vectors of solder joints from the standardized solder joint feature image sequence includes: Edge detection is performed on each frame of the standardized solder joint feature image sequence to extract the complete boundary of the solder joint contour; Calculate the area, perimeter, roundness, aspect ratio, and eccentricity of the weld joint profile, and combine these geometric parameters to form the geometric feature vector. The solder joint image is divided into several sub-regions, and the gray-level co-occurrence matrix is ​​calculated in each sub-region. Four texture feature indices—contrast, correlation, energy, and homogeneity—are extracted from the gray-level co-occurrence matrix. Statistical analysis is performed on the texture feature indices of all sub-regions to calculate the mean, variance, and extreme values, forming the texture distribution feature vector.

[0008] Further, the extraction of the heat conduction feature vector and cooling feature vector of the solder joint from the temperature distribution time-series curve includes: Identify the heating phase, peak plateau phase, and cooling phase in the temperature distribution time-series curve; During the heating phase, the initial slope, maximum slope, and average heating rate of the temperature rise are calculated to form a heating characteristic set. During the peak plateau phase, the stability index of the peak temperature and the plateau duration are calculated to form a peak characteristic set; During the cooling phase, the initial cooling rate, average cooling rate, and final steady-state temperature are calculated to form a cooling characteristic set. The heating characteristic group, peak characteristic group, and cooling characteristic group are combined in chronological order to form the heat conduction characteristic vector; Analyze the temperature drop curve during the cooling phase, fit an exponential decay model, extract the decay time constant and steady-state temperature deviation, and form the cooling feature vector.

[0009] Furthermore, the feature fusion of geometric shape feature vector, texture distribution feature vector, heat conduction feature vector, and cooling feature vector to construct a multi-dimensional solder joint quality feature descriptor includes: Normalize each geometric parameter in the geometric morphology feature vector to eliminate the influence of dimensions; Principal component analysis was performed on the texture indices in the texture distribution feature vector to extract the main texture feature components; Time-series encoding is performed on the features of each stage in the heat conduction feature vector to preserve temporal correlation; Logarithmic transformation is applied to the attenuation parameter in the cooling feature vector to enhance feature discriminative power. The feature-level fusion method is used to splice the processed geometric features, main texture feature components, time series encoded features and logarithmically transformed cooling features in a predetermined order; The concatenated feature vectors are subjected to dimensionality reduction processing, retaining the feature components with the most information content to form the multidimensional solder joint quality feature descriptor with fixed dimensions.

[0010] Furthermore, the step of inputting the multi-dimensional solder joint quality feature descriptor into a preset deep learning classification network and outputting the quality level label and defect type probability distribution for each solder joint includes: The deep learning classification network includes a feature extraction layer, an attention mechanism layer, and a fully connected classification layer; The feature extraction layer receives multi-dimensional solder joint quality feature descriptors and extracts deep feature representations through multiple convolutional and pooling layers. The attention mechanism layer analyzes the importance weights of different feature dimensions in the deep feature representation, enhances key features, and suppresses secondary features; The fully connected classification layer maps the weighted feature representations to a predefined classification space and outputs the probability value of each solder joint belonging to each quality level. The quality grade with the highest probability value is taken as the quality grade label; Meanwhile, the fully connected classification layer also outputs the probability distribution of solder joints belonging to various defect types, forming the defect type probability distribution, which includes incomplete soldering, cold soldering, bridging, insufficient solder, and excessive solder.

[0011] Furthermore, the step of locating the specific location of welding defects and analyzing the time stage of defect occurrence based on the probability distribution of defect types and the time series curve of temperature distribution includes: For defect types whose probability exceeds the threshold in the defect type probability distribution, backtrack to the original multimodal data set of the welding area; Locate the precise position of the defective solder joint in a standardized sequence of solder joint feature images and record its row and column coordinates; In the corresponding temperature distribution time-series curve, analyze the temperature anomalies of the weld joint during the welding process; By comparing the abnormal temperature points with the standard temperature curve of the welding process, the time periods in which the temperature deviated from the normal range can be identified. Based on the type and degree of temperature deviation, determine the specific process stage at which the defect may occur. The process stage includes the preheating stage, the welding stage, and the cooling stage. By combining information on the location of the weld joint and the stage at which the defect occurred, a complete result of defect localization and stage analysis is formed.

[0012] Furthermore, based on the quality grade label and defect location information of each weld point, a welding quality assessment report containing repair priorities and repair recommendations is generated, including: Statistically analyze the quality grade distribution of all weld joints, and calculate the proportion of qualified weld joints and the number of various defective weld joints; Based on the defect type and severity, each defective solder joint is assigned a repair priority score, with a higher priority score indicating that it needs to be repaired earlier. For each type of defect, retrieve the corresponding repair process parameters and key operating points from the repair knowledge base; Based on the location information of specific solder joints, a personalized repair path plan is generated to avoid affecting other solder joints during the repair process; The statistical results of quality grade distribution, the priority ranking of defective weld repairs, personalized repair suggestions, and repair path planning are integrated to form the aforementioned welding quality assessment report.

[0013] Furthermore, the comparison of abnormal temperature points with the standard temperature curve of the welding process to identify time periods in which the temperature deviates from the normal range includes: Retrieve the standard welding temperature curve corresponding to the current product model from the welding process parameter database. The standard welding temperature curve defines the ideal temperature range and duration for each process stage. Align the temperature distribution time-series curve of the weld point to be tested with the standard welding temperature curve on the time axis; Calculate the temperature difference between the two curves at each time point to form a temperature difference curve; Set a temperature difference threshold. When the temperature difference exceeds the threshold, mark that time point as a temperature anomaly. Analyze the distribution of continuous temperature anomalies and identify the duration of temperature anomalies; The specific stage at which the defect occurred is determined based on the duration of the temperature anomaly.

[0014] Furthermore, the step of retrieving corresponding repair process parameters and operational points from the repair knowledge base for each defect type includes: Establish a mapping table between defect types and repair processes. The mapping table records the repair methods, required tools, temperature parameters, and time parameters corresponding to different defect types. Based on the defect type identified in the defect type probability distribution, the corresponding repair method is found in the mapping table; Obtain detailed operating steps, safety precautions, and solutions to common problems for this repair method from the repair knowledge base; Based on the specific location characteristics of the solder joint and the layout of surrounding components, adjust the local parameters in the repair process parameters; Generate a personalized repair operation guide that includes a repair step sequence, parameter settings, tool selection, and risk warnings.

[0015] Compared with the prior art, the beneficial effects of the present invention are: By simultaneously acquiring multispectral scanning images and infrared thermal imaging time-series data of the welding area, and performing image enhancement and data registration, a spatiotemporally aligned multimodal data set is formed. This method overcomes the information limitations of conventional single optical detection, enabling the analysis of each weld point to simultaneously possess a standardized image sequence reflecting the surface material composition and microstructure, as well as a precise temperature change curve reflecting its complete thermal history. This fused data foundation makes it possible to comprehensively describe the characteristics of weld points from both static appearance and dynamic behavior dimensions.

[0016] From standardized temperature distribution time-series curves, heat conduction features characterizing the heat transfer rate and cooling features characterizing the solidification process were specifically extracted. These dynamic thermal features, fused with geometric and texture features extracted from images, construct a multidimensional weld joint quality feature descriptor that more fundamentally characterizes the physical process quality of welding. A deep learning classification network based on this descriptor can identify defect types that are not obvious in static images but exhibit abnormal thermal behavior patterns, improving the detection accuracy of latent defects such as poor heat wetting.

[0017] By correlating the probability distribution of defect types output by the network with the original time-series temperature distribution curve, the spatial location of defects can be pinpointed based on the time nodes of abnormal thermal patterns, and the process stage at which they occurred can be deduced. This allows the detection results to not only indicate "where and what kind of defects are present," but also reveal "when and what kind of process deviations, such as insufficient heating or excessively rapid cooling, may have caused the defects." This spatiotemporal location and causal correlation capability upgrades the output of quality inspection from a simple pass / fail judgment to a deep analysis report that includes diagnosis of the defect generation process. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the steps of the welding quality inspection method for flexible circuit boards in audio equipment according to the present invention. Figure 2 A flowchart for extracting geometric shapes and texture features; Figure 3 A graph showing the probability distribution of defect types and the attention weight analysis of features in response to welding defects in flexible circuit boards for audio equipment. Figure 4 A diagram showing the severity and repair priority of welding defects in flexible circuit boards for audio equipment; Figure 5 A comparative analysis of core temperature and time parameters for repairing welding defects in flexible circuit boards of audio equipment. Detailed Implementation

[0019] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0020] See Figure 1A method for inspecting the welding quality of flexible circuit boards for audio equipment is proposed. This method acquires multispectral scanning image data and infrared thermal imaging time-series data of the flexible circuit board surface of the audio equipment under test, forming a multimodal data set of the original welding area. Image enhancement and data registration are performed on this original welding area multimodal data set to generate a standardized sequence of solder joint feature images and corresponding temperature distribution time-series curves. Geometric morphology feature vectors and texture distribution feature vectors of the solder joints are extracted from the standardized solder joint feature image sequence, and heat conduction feature vectors and cooling feature vectors of the solder joints are extracted from the temperature distribution time-series curves. These feature vectors are then fused to construct a multidimensional solder joint quality feature descriptor. This multidimensional solder joint quality feature descriptor is input into a pre-defined deep learning classification network. The deep learning classification network outputs a quality level label and defect type probability distribution for each solder joint. Based on the defect type probability distribution and temperature distribution time-series curves, the specific location of the welding defect is located, and the time stage of defect occurrence is analyzed. Based on the quality level label and defect location information of each solder joint, a welding quality assessment report containing repair priorities and repair suggestions is generated.

[0021] In one embodiment of the present invention, see [reference] Figure 2 The process involves denoising the multispectral scan image data, employing an adaptive filtering algorithm to eliminate image noise and enhance the contrast between solder joints and the background. Temperature calibration is performed on the infrared thermal imaging time-series data, converting the original thermal imaging grayscale values ​​into precise temperature values, establishing a time-temperature correspondence, identifying the solder joint contour positions in the multispectral images, establishing a local coordinate system with the solder joint center as the reference, and mapping the infrared thermal imaging data to the local coordinate system to ensure precise correspondence between the positions of each solder joint in the thermal imaging data and the multispectral images. Size normalization is performed on the registered multispectral image sequence to adjust all solder joint images to a uniform resolution, forming a standardized solder joint feature image sequence. Temperature change data for each solder joint in the registered thermal imaging data is extracted and arranged chronologically to form the corresponding temperature distribution time-series curve. Edge detection is performed on each frame of the standardized solder joint feature image sequence to extract the complete boundary of the solder joint contour. The area, perimeter, roundness, aspect ratio, and eccentricity of the solder joint contour are calculated. These geometric parameters are combined to form a geometric morphology feature vector. The solder joint image is divided into several sub-regions. The gray-level co-occurrence matrix is ​​calculated in each sub-region. Four texture feature indicators, namely contrast, correlation, energy, and homogeneity, are extracted from the gray-level co-occurrence matrix. Statistical analysis is performed on the texture feature indicators of all sub-regions to calculate the mean, variance, and extreme values, forming a texture distribution feature vector.

[0022] In practical implementation, the image enhancement, data registration, and feature extraction processes for the flexible circuit board welding quality inspection method for audio equipment are operated as follows: Denoising of the multispectral scan image data is performed using an adaptive filtering algorithm to eliminate image noise. One optional adaptive filtering algorithm adjusts the filtering coefficients based on local image variance. This can be expressed as a weighted average of pixel grayscale values ​​within the filtering window, with the weights inversely proportional to the grayscale difference between the pixel and the center pixel. This process enhances the contrast between the solder joint and the background. Temperature calibration is performed on the infrared thermal imaging time-series data. Based on the blackbody radiation law, the original thermal imaging grayscale values ​​are converted into precise temperature values, establishing a time-temperature correspondence. The solder joint contour positions in the multispectral images are identified using edge detection and contour fitting algorithms, establishing a local coordinate system with the solder joint center as the reference. Affine transformation is used to map the infrared thermal imaging data to the local coordinate system, ensuring that the position of each solder joint in the thermal imaging data and multispectral image corresponds precisely. Size normalization is performed on the registered multispectral image sequence using a bilinear interpolation algorithm to adjust all solder joint images to a uniform resolution, forming a standardized solder joint feature image sequence. The temperature change data of each solder joint in the registered thermal imaging data is extracted and arranged in chronological order to form the corresponding temperature distribution time series curve.

[0023] In some embodiments, the process of extracting geometric morphological feature vectors and texture distribution feature vectors of solder joints from a standardized sequence of solder joint feature images is specifically implemented. Edge detection is performed on each frame of the standardized solder joint feature image sequence, and the Canny operator is used to extract the complete boundary of the solder joint contour. The area, perimeter, roundness, aspect ratio, and eccentricity of the solder joint contour are calculated, and these geometric parameters are combined to form a geometric morphological feature vector. The solder joint image is divided into several sub-regions using a uniform grid method. A gray-level co-occurrence matrix is ​​calculated within each sub-region. The gray-level co-occurrence matrix describes the probability of occurrence of pixel pairs with specific spatial relationships. Four texture feature indices—contrast, correlation, energy, and homogeneity—are extracted from the gray-level co-occurrence matrix. Contrast reflects the image's sharpness and the depth of texture grooves; correlation measures the directionality of the image texture; energy indicates the uniformity of the image texture; and homogeneity reflects the uniformity of local areas of the image.

[0024] It is understandable that the calculation of texture feature indicators involves mathematical operations. The contrast calculation formula can be expressed as:

[0025] in Indicates the contrast value. It is the number of gray levels in the image. and These are the gray levels of a pixel pair. Given a spatial distance and direction Lower grayscale pairs The probability of occurrence in the gray-level co-occurrence matrix. Statistical analysis is performed on the texture feature indices of all sub-regions to calculate the mean, variance, and extreme values, forming a texture distribution feature vector.

[0026] In some embodiments, the calculation of the geometric parameters of the solder joint profile has a clear mathematical definition. The area is the total number of pixels inside the profile, the perimeter is the total length of pixels at the profile boundary, and the roundness is determined by the formula... The aspect ratio is calculated as the ratio of the longer side to the shorter side of the smallest bounding rectangle of the profile, and the eccentricity is the distance between the two foci after fitting the profile ellipse. These parameters collectively constitute the geometric feature vector describing the two-dimensional shape of the weld joint.

[0027] Optionally, after establishing a local coordinate system, the mapping accuracy of the infrared thermal imaging data is evaluated. The evaluation method includes calculating the coordinate deviation between the center coordinates of the solder joint in the multispectral image and the corresponding point in the mapped thermal imaging data. When the average coordinate deviation exceeds a preset pixel threshold, the affine transformation parameter calculation process is re-executed until the accuracy meets the requirements. Optionally, the uniform resolution used for size normalization is determined comprehensively based on the average physical size of the solder joint and the resolution of the imaging system, ensuring that the normalized image retains key details of the solder joint.

[0028] In one embodiment of the present invention, a heating phase, a peak plateau phase, and a cooling phase are identified in the temperature distribution time series curve. In the heating phase, the initial slope, maximum slope, and average heating rate of the temperature rise are calculated to form a heating feature group. In the peak plateau phase, the stability index of the peak temperature and the plateau duration are calculated to form a peak feature group. In the cooling phase, the initial cooling rate, average cooling rate, and final stable temperature of the temperature drop are calculated to form a cooling feature group. The heating feature group, peak feature group, and cooling feature group are combined in chronological order to form a heat conduction feature vector. The temperature drop curve of the cooling phase is analyzed, an exponential decay model is fitted, and the decay time constant and steady-state temperature deviation are extracted to form a cooling feature vector. The geometric parameters in the geometric morphology feature vector are normalized to eliminate the influence of dimensions. Principal component analysis is performed on the texture index in the texture distribution feature vector to extract the main texture feature components. Time series encoding is performed on the features of each stage in the heat conduction feature vector to retain temporal correlation. Logarithmic transformation is performed on the attenuation parameter in the cooling feature vector to enhance feature discriminability. A feature-level fusion method is adopted to concatenate the processed geometric morphology features, main texture feature components, time series encoded features, and logarithmically transformed cooling features in a predetermined order. The concatenated feature vector is then dimensionality reduced to retain the feature component with the most information, forming a fixed-dimensional multidimensional solder joint quality feature descriptor.

[0029] In practical implementation, feature extraction of the temperature distribution time series curve and construction of multi-dimensional solder joint quality feature descriptors are carried out as follows: Identification of the heating phase, peak plateau phase, and cooling phase in the temperature distribution time series curve is based on the analysis of the first derivative of the temperature curve. Phase division is achieved by identifying inflection points where the rate of temperature change changes from positive to zero and from negative to zero. In the heating phase, the initial slope, maximum slope, and average heating rate are calculated. The initial slope is the linear fitting slope between the starting point of the temperature curve and the first significant inflection point. The maximum slope is the maximum instantaneous slope among all sampling points within the heating phase. The average heating rate is the ratio of the temperature change to the time change throughout the heating phase. These parameters form the heating feature set. In the peak plateau phase, the stability index of the peak temperature and the plateau duration are calculated. The stability index is the reciprocal of the standard deviation of the plateau temperature, and the plateau duration is the length of time the temperature remains within a preset percentage range of the peak value. These parameters form the peak feature set. During the cooling phase, the initial cooling rate, average cooling rate, and final steady-state temperature are calculated. The initial cooling rate is the instantaneous slope from the start of the cooling phase to the first sampling point. The average cooling rate is the ratio of the temperature change to the time change throughout the cooling phase. The final steady-state temperature is the average temperature of multiple sampling points at the end of the cooling curve. These parameters form a cooling characteristic set. The heating characteristic set, peak characteristic set, and cooling characteristic set are combined in chronological order to form a heat conduction characteristic vector. The temperature decrease curve during the cooling phase is analyzed, an exponential decay model is fitted, and the decay time constant and steady-state temperature deviation are extracted to form the cooling characteristic vector.

[0030] In some embodiments, the extraction of cooling feature vectors involves fitting a mathematical model. The fitted exponential decay model employs a nonlinear least squares method, and the model can be expressed as follows: ,in Indicates the time during the cooling phase. Temperature value at that time This represents the theoretical steady-state temperature obtained from the fitting. This indicates the peak temperature at the start of the cooling phase. This represents the decay time constant obtained from the fitting. It is the base of the natural logarithm. The decay time constant is obtained directly from the fitting results. The steady-state temperature deviation is calculated by comparing the final measured steady-state temperature with the fitted theoretical steady-state temperature. The absolute difference is obtained. Decay time constant. Together with the steady-state temperature deviation, they constitute the cooling characteristic vector.

[0031] It is understandable that the construction of heat conduction and cooling feature vectors provides input for subsequent feature fusion. Normalization of each geometric parameter in the geometric morphology feature vector is performed using a minimum-maximum scaling method, linearly transforming each geometric parameter to the range of zero to one, eliminating the influence of dimensions. Principal component analysis is performed on the texture indices in the texture distribution feature vector, calculating the eigenvalues ​​and eigenvectors of the texture index covariance matrix, and selecting the top few principal components whose cumulative contribution rate exceeds a set threshold as the main texture feature components. Time-series encoding is performed on the features of each stage in the heat conduction feature vector, using one-hot encoding to represent the process stage to which the feature belongs, preserving temporal correlation. Logarithmic transformation is performed on the decay parameters in the cooling feature vector, i.e., calculating the base-10 logarithmic values ​​for the decay time constant and steady-state temperature deviation, enhancing feature discriminability.

[0032] In some embodiments, the feature fusion and dimensionality reduction process is specifically implemented. A feature-level fusion method is used to concatenate the processed geometric morphology features, main texture feature components, time-series encoded features, and logarithmically transformed cooling features into a high-dimensional temporary feature vector in a predetermined order. The predetermined order is based on the order of feature origin, namely: geometric morphology features, principal components of texture distribution features, time-series encoded heat conduction features, and logarithmically transformed cooling features. Dimensionality reduction of the concatenated temporary feature vector is performed using linear discriminant analysis to find the projection direction that maximizes the inter-class divergence and minimizes the intra-class divergence among different types of solder joint features. The top few feature components with the highest information content after projection are retained to form a fixed-dimensional multidimensional solder joint quality feature descriptor.

[0033] Optionally, the threshold for the cumulative contribution rate in principal component analysis can be adjusted according to the requirements for the degree of feature information retention in specific applications. Optionally, the number of feature components retained after dimensionality reduction by linear discriminant analysis is determined by cross-validation, which can achieve balanced recognition performance in subsequent classification tasks. It can be understood that the steps of normalization, principal component analysis, time series encoding, logarithmic transformation, and linear discriminant analysis work together to transform the original features from different modalities and dimensions into a compact and discriminative multidimensional solder joint quality feature descriptor.

[0034] In one embodiment of the present invention, the deep learning classification network includes a feature extraction layer, an attention mechanism layer, and a fully connected classification layer. The feature extraction layer receives multi-dimensional solder joint quality feature descriptors and extracts deep feature representations through multiple convolutional and pooling layers. The attention mechanism layer analyzes the importance weights of different feature dimensions in the deep feature representations, enhances key features, and suppresses secondary features. The fully connected classification layer maps the weighted feature representations to a predefined classification space and outputs the probability value of each solder joint belonging to each quality level. The quality level with the highest probability value is taken as the quality level label. At the same time, the fully connected classification layer also outputs the probability distribution of solder joints belonging to various defect types, forming a defect type probability distribution. Defect types include cold solder joints, bridging, insufficient solder, and excessive solder. For defect types whose probability exceeds a threshold in the defect type probability distribution, the original welding area multimodal dataset is traced back to locate the precise location of the defective weld point in the standardized weld point feature image sequence, and its row and column coordinates are recorded. In the corresponding temperature distribution time series curve, the temperature anomalies of the weld point during the welding process are analyzed. The temperature anomalies are compared with the standard temperature curve of the welding process to identify the time period when the temperature deviates from the normal range. Based on the type and degree of temperature deviation, the specific process stage at which the defect may occur is determined. The process stages include the preheating stage, the welding stage, and the cooling stage. By combining the weld point location information and the defect generation stage information, a complete defect location and stage analysis result is formed.

[0035] In practice, the process of inputting the multi-dimensional solder joint quality feature descriptor into a predefined deep learning classification network and obtaining its output is implemented as follows: The deep learning classification network includes a feature extraction layer, an attention mechanism layer, and a fully connected classification layer. The feature extraction layer receives the multi-dimensional solder joint quality feature descriptor and extracts deep feature representations through multiple convolutional and pooling layers. The convolutional layers use one-dimensional convolutional kernels to perform sliding calculations along the feature dimensions, and the pooling layers use max pooling to downsample the convolutional output. The attention mechanism layer analyzes the importance weights of different feature dimensions in the deep feature representation, enhancing key features and suppressing secondary features. The fully connected classification layer maps the weighted feature representation to a predefined classification space, outputting the probability value of each solder joint belonging to each quality level, and taking the quality level with the highest probability value as the quality level label. Simultaneously, the fully connected classification layer also outputs the probability distribution of solder joints belonging to various defect types, forming a defect type probability distribution. Defect types include incomplete soldering, cold soldering, bridging, insufficient solder, and excessive solder.

[0036] In some embodiments, the specific operation of the attention mechanism layer involves weight calculation. The attention mechanism layer first calculates the importance score for each feature dimension in the deep feature representation. The importance score is obtained through a small, trainable fully connected network. The importance scores of all feature dimensions are then input into a softmax function for normalization, yielding the attention weight for each feature dimension. The attention weight is then multiplied element-wise with the original deep feature representation to achieve feature enhancement and suppression. The calculation process of the attention weight can be formally represented as follows:

[0037] in Indicates the first Attention weights for each feature dimension. Indicates the first Importance scores for each feature dimension This represents the total number of dimensions in the deep feature representation. This represents the natural exponential function. It can be understood through attention weights... By adjusting the settings, deep learning classification networks can focus on feature dimensions that contribute more to the classification task.

[0038] In practice, the process of locating welding defects and analyzing time stages based on defect type probability distribution and temperature distribution time series curves is as follows: For defect types with a probability exceeding a preset threshold in the defect type probability distribution, the original multimodal data set of the welding area is retrieved. The precise location of the weld point with the defect is located in the standardized weld point feature image sequence, and its row and column coordinates are recorded. In the corresponding temperature distribution time series curve, the temperature anomalies of the weld point during the welding process are analyzed. The temperature anomalies are compared with the standard temperature curve of the welding process to identify the time periods when the temperature deviates from the normal range. Based on the type and degree of temperature deviation, the specific process stage at which the defect may occur is determined, including the preheating stage, welding stage, and cooling stage. By combining the weld point location information and the defect occurrence stage information, a complete defect location and stage analysis result is formed.

[0039] In some embodiments, the analysis process for temperature anomalies is specifically implemented. Analyzing temperature anomalies during the welding process begins by calculating the first-order difference sequence of the temperature distribution time-series curve, identifying time points where the difference value exceeds a preset positive or negative threshold as candidate anomalies. These candidate anomalies undergo a continuous test, retaining only time points that consistently exhibit an abnormal trend across multiple consecutive sampling periods; these are then confirmed as temperature anomalies. When comparing temperature anomalies with the standard temperature curve of the welding process, the standard temperature curve is obtained from the process parameter database in the form of a time-temperature lookup table. The period of temperature deviation from the normal range is identified by comparing the actual temperature value at each temperature anomaly time point with the corresponding standard temperature value on the standard temperature curve. When the absolute difference between the actual temperature value and the standard temperature value consistently exceeds the tolerance range, this period is determined to be a time period of temperature deviation from the normal range.

[0040] Optionally, the preset threshold is set based on the welding process requirements of the specific product model and the statistical results of historical inspection data. Optionally, the tolerance range of the standard temperature curve is determined based on the thermal stability index of the welding equipment and the heat capacity characteristics of the material. It can be understood that by associating the probability output of the deep learning classification network with the original spatiotemporal data, the abstract defect classification probability can be transformed into specific physical location and process time information, thereby completing the transition from detection to preliminary diagnosis.

[0041] See Figure 3 This is a graph analyzing the probability distribution and feature attention weights of welding defects in flexible circuit boards for audio equipment. It represents a core visualization result of the "deep learning classification network output stage" in the welding quality inspection process. The predicted probability of a cold solder joint is approximately 0.89, significantly higher than the probability threshold of 0.5, making it the main defect type detected. The feature attention weight for a cold solder joint is approximately 0.91, the highest among all defects, indicating that the deep learning network highly focuses on feature dimensions related to cold solder joints during classification, validating the effectiveness of the features. The predicted probabilities of cold solder joints, bridging, insufficient solder, and excessive solder are all far below the 0.5 threshold, indicating that these defects have an extremely low probability of occurrence in this detection. The feature attention weight for insufficient solder is approximately 0.85, the second highest value, indicating strong discriminative power of its corresponding feature. The attention weights for cold solder joints, bridging, and excessive solder are in the range of 0.6 to 0.8, reflecting a relatively balanced contribution of these defect features.

[0042] In one embodiment of the present invention, the quality grade distribution of all weld points is statistically analyzed, the proportion of qualified weld points and the number of various defective weld points are calculated, and a repair priority score is assigned to each defective weld point according to the defect type and severity. The higher the priority score, the earlier the repair is required. For each defect type, the corresponding repair process parameters and operation points are retrieved from the repair knowledge base. Combined with the location information of the specific weld point, a personalized repair path plan is generated to avoid affecting other weld points during the repair process. The statistical results of the quality grade distribution, the repair priority ranking of defective weld points, personalized repair suggestions, and repair path plans are integrated to form a structured welding quality assessment report. The standard welding temperature curve corresponding to the current product model is retrieved from the welding process parameter database. The standard welding temperature curve defines the ideal temperature range and duration of each process stage. The temperature distribution time sequence curve of the weld point to be tested is aligned with the standard welding temperature curve on the time axis, and the temperature difference between the two curves at each time point is calculated to form a temperature difference curve. A temperature difference threshold is set. When the temperature difference exceeds the threshold, the time point is marked as a temperature anomaly point. The distribution of continuous temperature anomalies is analyzed, the duration of temperature anomalies is identified, and the specific stage of defect occurrence is determined according to the process stage in which the temperature anomaly duration is located.

[0043] In practical implementation, feature extraction of the temperature distribution time series curve and construction of multi-dimensional solder joint quality feature descriptors are carried out as follows: Identification of the heating phase, peak plateau phase, and cooling phase in the temperature distribution time series curve is based on the analysis of the first derivative of the temperature curve. Phase division is achieved by identifying inflection points where the rate of temperature change changes from positive to zero and from negative to zero. In the heating phase, the initial slope, maximum slope, and average heating rate are calculated. The initial slope is the linear fitting slope between the starting point of the temperature curve and the first significant inflection point. The maximum slope is the maximum instantaneous slope among all sampling points within the heating phase. The average heating rate is the ratio of the temperature change to the time change throughout the heating phase. These parameters form the heating feature set. In the peak plateau phase, the stability index of the peak temperature and the plateau duration are calculated. The stability index is the reciprocal of the standard deviation of the plateau temperature, and the plateau duration is the length of time the temperature remains within a preset percentage range of the peak value. These parameters form the peak feature set. During the cooling phase, the initial cooling rate, average cooling rate, and final steady-state temperature are calculated. The initial cooling rate is the instantaneous slope from the start of the cooling phase to the first sampling point. The average cooling rate is the ratio of the temperature change to the time change throughout the cooling phase. The final steady-state temperature is the average temperature of multiple sampling points at the end of the cooling curve. These parameters form a cooling characteristic set. The heating characteristic set, peak characteristic set, and cooling characteristic set are combined in chronological order to form a heat conduction characteristic vector. The temperature decrease curve during the cooling phase is analyzed, an exponential decay model is fitted, and the decay time constant and steady-state temperature deviation are extracted to form the cooling characteristic vector.

[0044] In practice, the process of generating a welding quality assessment report is as follows: The quality grade distribution of all weld points is statistically analyzed, the proportion of qualified weld points and the number of various defective weld points are calculated, and a repair priority score is assigned to each defective weld point based on its defect type and severity. A higher priority score indicates that repair should be carried out earlier. For each defect type, corresponding repair process parameters and operational points are retrieved from the repair knowledge base. Combined with the specific weld point's location information, a personalized repair path plan is generated to avoid affecting other weld points during the repair process. The statistical results of the quality grade distribution, the repair priority ranking of defective weld points, personalized repair suggestions, and the repair path plan are integrated to form a structured welding quality assessment report. The repair priority score can be calculated based on a scoring function, with the following form:

[0045] in Indicates the repair priority score. This represents a base score indicating the severity of this defect type based on historical data. This represents the urgency coefficient derived from the time stage at which the defect occurred. and These are the severity weighting and urgency weighting coefficients, respectively. Defect type and corresponding base severity score. See Table 1: Table 1: Defect Type and Repair Priority Scoring Mapping Table

[0046] In some embodiments, the operation of comparing temperature anomalies with the standard temperature curve of the welding process and identifying time periods involves retrieving the standard welding temperature curve corresponding to the current product model from the welding process parameter database. The standard welding temperature curve defines the ideal temperature range and duration for each process stage. The temperature distribution time sequence curve of the weld point to be inspected is aligned with the standard welding temperature curve on the time axis, with the alignment operation based on the welding process's start trigger signal as a common zero point. The temperature difference between the two curves at each time point is calculated to form a temperature difference curve. A temperature difference threshold is set, and when the temperature difference exceeds the threshold, that time point is marked as a temperature anomaly. The distribution of continuous temperature anomalies is analyzed to identify the duration of temperature anomalies. Based on the process stage in which the temperature anomaly duration occurs, the specific stage at which the defect occurs is determined.

[0047] It is understandable that generating the temperature difference curve involves point-by-point calculations for each sampling time. Temperature difference The calculation method is as follows: | in: The temperature of the solder joint to be tested. The standard temperature is K, representing the total number of sampling points. The logic for identifying the duration of temperature anomalies is as follows: when the temperature exceeds a certain threshold for an anomaly... sampling points When all values ​​exceed the temperature difference threshold, the time interval covered by these consecutive points is determined to be a temperature anomaly duration. Parameters The value of is related to the sampling frequency and is used to filter out short-term random temperature fluctuations.

[0048] In some embodiments, the process of generating repair path planning is specifically implemented. Generating personalized repair path planning requires combining the location information of specific solder joints. The location information is represented by the row and column coordinates of the solder joints on the flexible circuit board. The path planning algorithm takes the starting point of the repair device as the origin and calculates an access sequence with the shortest total travel distance based on the coordinates of all solder joints to be repaired. This sequence is also scheduled in descending order of repair priority score. For solder joints with the same priority score, the order is arranged according to their natural proximity relationship in physical location. The generated repair path planning is output in the form of a sequence of coordinate points, with the solder joint identifier and priority score corresponding to each coordinate point.

[0049] Optionally, the temperature difference threshold is not a fixed value, but rather proportional to the ideal temperature range of the current process stage in the standard welding temperature profile. A relatively larger threshold is used in process stages with drastic temperature changes, while a relatively strict threshold is used during temperature plateau periods. Optionally, when retrieving repair process parameters from the repair knowledge base, the repair knowledge base exists in the form of a structured database. Through the defect type index, users can find text-formatted operation points, parameter lists, and necessary safety warnings.

[0050] Understandably, the statistical results of the quality grade distribution are presented in the form of percentages and counts, the repair priority ranking lists all defective weld points and their scores in the form of a list, the personalized repair suggestions include operation guidelines retrieved from the knowledge base and adapted to specific locations, and the repair path planning is given in the form of an ordered list or graphical coordinates. These components are integrated into a unified document template to form the final welding quality assessment report.

[0051] See Figure 4 This is a defect severity and repair priority analysis chart for welding defects in flexible circuit boards of audio equipment. It is a core visualization result of the "quality assessment and repair planning stage" in the welding quality inspection process. This result directly corresponds to the core logic in the patent embodiment. The chart quantifies the severity, number, and repair priority of different defects, providing data support for generating a welding quality assessment report. The repair priority score is a weighted result of the combined severity and number of defects, which can be directly used to guide the allocation and sequencing of resources for on-site repairs, avoiding blind operations. The high number and high priority of cold solder joints directly point to potential process problems such as "insufficient heat conduction" or "insufficient solder wettability" in the welding process, providing a clear entry point for process optimization.

[0052] In one embodiment of the present invention, a mapping table between defect types and repair processes is established. The mapping table records the repair methods, required tools, temperature parameters, and time parameters corresponding to different defect types. Based on the defect type identified in the defect type probability distribution, the corresponding repair method is searched in the mapping table. Detailed operation steps, safety precautions, and solutions to common problems of the repair method are obtained from the repair knowledge base. Combining the location characteristics of the specific solder joint and the layout of surrounding components, the local parameters in the repair process parameters are adjusted to generate a personalized repair operation guide that includes a repair step sequence, parameter settings, tool selection, and risk warnings.

[0053] In practice, for each defect type, the corresponding repair process parameters and key operation points are retrieved from the repair knowledge base and executed as follows: A mapping table between defect types and repair processes is established. The mapping table records the repair methods, required tools, temperature parameters, and time parameters corresponding to different defect types. Based on the defect type identified in the defect type probability distribution, the corresponding repair method is found in the mapping table. Detailed operation steps, safety precautions, and solutions to common problems for the repair method are obtained from the repair knowledge base. Combining the location characteristics of the specific solder joint and the layout of surrounding components, the local parameters in the repair process parameters are adjusted to generate a personalized repair operation guide that includes a repair step sequence, parameter settings, tool selection, and risk warnings.

[0054] In some embodiments, the process of establishing a mapping relationship table between defect types and repair processes is specifically implemented. The mapping relationship table is constructed in the form of a database table. The columns of the table include at least the defect type identifier, repair method name, main tool list, core temperature parameter, key time parameter, and associated knowledge base entry number. The defect type identifier corresponds one-to-one with the defect type name output by the deep learning classification network. The repair method name describes the specific process, such as "hot air gun resoldering", "solder wire cleaning", "dispensing soldering", etc. The main tool list lists the tools and equipment necessary to perform the method, such as constant temperature soldering iron, hot air workstation, and microscope. The core temperature parameter indicates the target temperature or temperature range of the heating tool during operation. The key time parameter indicates the duration of heating or cooling waiting time. The associated knowledge base entry number is linked as a foreign key to the repair knowledge base table storing detailed text information.

[0055] It is understandable that adjusting local parameters involves recalculating the standard repair parameters. The adjustment process is based on the characteristics of the solder joint location and the specific layout of surrounding components. An optional formula for adjusting local parameters can be expressed as:

[0056] in This represents the adjusted local parameter value. This represents the basic parameter value obtained from the mapping table. This represents the adjustment factor determined based on component layout density and thermal sensitivity. This indicates a correction factor determined by the specific location characteristics of the solder joint. Location characteristics include whether the solder joint is located at the edge of the board, whether it is close to large heat sinks or plastic connectors, and surrounding component layout information, including the minimum spacing between adjacent components and the presence of heat-sensitive devices.

[0057] In practice, the process of generating personalized repair operation guidelines follows a structured template. The guidelines include sections such as "Defect Overview," "Pre-Repair Preparation," "Step-by-Step Operation Procedure," "Parameter Monitoring Points," and "Post-Repair Inspection." The Defect Overview section references specific types and confidence levels from the probability distribution of defect types. The Pre-Repair Preparation section lists the necessary tools obtained from the mapping table and prompts for safety equipment checks. The Step-by-Step Operation Procedure section integrates detailed operation steps from the repair knowledge base with adjusted local parameter values. The Parameter Monitoring Points section specifically marks the key settings and allowable tolerances for temperature and time parameters. The Post-Repair Inspection section provides suggested methods for visual inspection and power-on / off testing.

[0058] In some embodiments, the repair knowledge base is constructed using a relational database combined with document storage. The relational database stores structured mapping tables and parameter indexes, while the document storage system stores unstructured detailed operation steps, safety precautions, diagrams, and solutions to common problems. When the system finds the corresponding repair method and associated knowledge base entry number in the mapping table based on the defect type, it concurrently retrieves all relevant text and multimedia content from the document storage system using that number and assembles them.

[0059] Optional, adjustment coefficient The determination is based on a predefined set of rules, which considers factors such as the average spacing of surrounding components, the presence of devices below a specific temperature threshold, and the number of layers and copper thickness of the flexible circuit board in that area. Optional, a correction factor... The value is obtained by querying a pre-configured location feature weight table, which assigns specific values ​​to different combinations of location features.

[0060] Understandably, the final personalized repair operation guide is a comprehensive document that integrates defect information, standard processes, local adaptation parameters, and safety tips. This document is output as a core component of the welding quality assessment report and is used to directly guide the repair operations on the production line or repair site.

[0061] See Figure 5This is a comparative analysis chart of core temperature and time parameters for repairing soldering defects on flexible circuit boards for audio equipment. It represents a core visualization result from the "repair process parameter retrieval and optimization stage" of the soldering quality inspection process. The adjusted temperatures for all defect types are significantly lower than the base temperature, verifying the effectiveness of the "local parameter adjustment formula" in the patent. This formula conservatively corrects the base parameters using layout coefficients and position factors to adapt to the thermally sensitive characteristics of flexible circuit boards. The bridging defect has the highest base temperature (approximately 390℃), adjusted temperature (approximately 300℃), and adjusted time (approximately 7.5s), indicating that repairing this defect requires higher heat input and a longer operation time, making it the most complex defect type. The insufficient solder defect has the shortest adjusted time (approximately 5.2s), and its adjusted temperature (approximately 365℃) is only slightly lower than the base temperature, indicating that the repair process for this defect is the simplest and requires the lowest heat input.

[0062] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for inspecting the welding quality of flexible circuit boards in audio equipment, characterized in that, Includes the following steps: Acquire multispectral scanning image data and infrared thermal imaging time-series data of the flexible circuit board surface of the audio device under test to form a set of original welding area multimodal data; Image enhancement and data registration are performed on the original welding area multimodal data set to generate a standardized weld point feature image sequence and corresponding temperature distribution time-series curve; Extract the geometric morphology feature vector and texture distribution feature vector of the solder joint from the standardized solder joint feature image sequence, and extract the heat conduction feature vector and cooling feature vector of the solder joint from the temperature distribution time series curve; The geometric shape feature vector, texture distribution feature vector, heat conduction feature vector and cooling feature vector are fused to construct a multi-dimensional solder joint quality feature descriptor. The multidimensional weld joint quality feature descriptor is input into a preset deep learning classification network, and the deep learning classification network outputs the quality level label and defect type probability distribution for each weld joint. Based on the probability distribution of the defect type and the time series curve of the temperature distribution, the specific location of the welding defect is located and the time stage of defect occurrence is analyzed. Based on the quality grade label and defect location information of each weld point, a welding quality assessment report containing repair priorities and repair recommendations is generated.

2. The method for inspecting the welding quality of flexible circuit boards for audio equipment according to claim 1, characterized in that, The process of image enhancement and data registration of the original welding area multimodal data set to generate a standardized weld point feature image sequence and corresponding temperature distribution time-series curve includes: Denoising processing is performed on the multispectral scan image data, and an adaptive filtering algorithm is used to eliminate image noise and enhance the contrast between the solder joints and the background. Temperature calibration is performed on the time-series data of infrared thermal imaging, converting the original thermal imaging grayscale values ​​into precise temperature values ​​and establishing a time-temperature correspondence. Identify the contour location of solder joints in multispectral images and establish a local coordinate system with the center of the solder joint as the reference. Infrared thermal imaging data is mapped to the local coordinate system to ensure that the position of each solder joint corresponds precisely in the thermal imaging data and multispectral image; The registered multispectral image sequence is subjected to size normalization processing to adjust all solder joint images to a uniform resolution, forming the standardized solder joint feature image sequence. Temperature change data for each solder joint is extracted from the registered thermal imaging data and arranged in chronological order to form the corresponding temperature distribution time-series curve.

3. The method for inspecting the welding quality of flexible circuit boards for audio equipment according to claim 2, characterized in that, The extraction of geometric morphological feature vectors and texture distribution feature vectors of solder joints from a standardized sequence of solder joint feature images includes: Edge detection is performed on each frame of the standardized solder joint feature image sequence to extract the complete boundary of the solder joint contour; Calculate the area, perimeter, roundness, aspect ratio, and eccentricity of the weld joint profile, and combine these geometric parameters to form the geometric feature vector. The solder joint image is divided into several sub-regions, and the gray-level co-occurrence matrix is ​​calculated in each sub-region. Four texture feature indices—contrast, correlation, energy, and homogeneity—are extracted from the gray-level co-occurrence matrix. Statistical analysis is performed on the texture feature indices of all sub-regions to calculate the mean, variance, and extreme values, forming the texture distribution feature vector.

4. The method for inspecting the welding quality of flexible circuit boards for audio equipment according to claim 2, characterized in that, The extraction of the heat conduction feature vector and cooling feature vector of the solder joint from the temperature distribution time-series curve includes: Identify the heating phase, peak plateau phase, and cooling phase in the temperature distribution time-series curve; During the heating phase, the initial slope, maximum slope, and average heating rate of the temperature rise are calculated to form a heating characteristic set. During the peak plateau phase, the stability index of the peak temperature and the plateau duration are calculated to form a peak characteristic set; During the cooling phase, the initial cooling rate, average cooling rate, and final steady-state temperature are calculated to form a cooling characteristic set. The heating characteristic group, peak characteristic group, and cooling characteristic group are combined in chronological order to form the heat conduction characteristic vector; Analyze the temperature drop curve during the cooling phase, fit an exponential decay model, extract the decay time constant and steady-state temperature deviation, and form the cooling feature vector.

5. The method for inspecting the welding quality of flexible circuit boards for audio equipment according to claim 1, characterized in that, The process of fusing geometric shape feature vectors, texture distribution feature vectors, heat conduction feature vectors, and cooling feature vectors to construct a multi-dimensional solder joint quality feature descriptor includes: Normalize each geometric parameter in the geometric morphology feature vector to eliminate the influence of dimensions; Principal component analysis was performed on the texture indices in the texture distribution feature vector to extract the main texture feature components. Time-series encoding is performed on the features of each stage in the heat conduction feature vector to preserve temporal correlation; Logarithmic transformation is applied to the attenuation parameter in the cooling feature vector to enhance feature discriminative power. The feature-level fusion method is used to splice the processed geometric features, main texture feature components, time series encoded features and logarithmically transformed cooling features in a predetermined order; The concatenated feature vectors are subjected to dimensionality reduction processing, retaining the feature components with the most information content to form the multidimensional solder joint quality feature descriptor with fixed dimensions.

6. The method for inspecting the welding quality of flexible circuit boards for audio equipment according to claim 1, characterized in that, The process of inputting multi-dimensional solder joint quality feature descriptors into a preset deep learning classification network and outputting a quality level label and defect type probability distribution for each solder joint includes: The deep learning classification network includes a feature extraction layer, an attention mechanism layer, and a fully connected classification layer; The feature extraction layer receives multidimensional solder joint quality feature descriptors and extracts deep feature representations through multiple convolutional and pooling layers. The attention mechanism layer analyzes the importance weights of different feature dimensions in the deep feature representation, enhances key features, and suppresses secondary features; The fully connected classification layer maps the weighted feature representations to a predefined classification space and outputs the probability value of each solder joint belonging to each quality level. The quality grade with the highest probability value is selected as the quality grade label. Meanwhile, the fully connected classification layer also outputs the probability distribution of solder joints belonging to various defect types, forming the defect type probability distribution, which includes incomplete soldering, cold soldering, bridging, insufficient solder, and excessive solder.

7. The method for inspecting the welding quality of flexible circuit boards for audio equipment according to claim 6, characterized in that, The method of locating the specific location of welding defects and analyzing the time stage of defect occurrence based on the probability distribution of defect types and the time series curve of temperature distribution includes: For defect types whose probability exceeds the threshold in the defect type probability distribution, backtrack to the original multimodal data set of the welding area; Locate the precise position of the defective solder joint in a standardized sequence of solder joint feature images and record its row and column coordinates; In the corresponding temperature distribution time-series curve, analyze the temperature anomalies of the weld joint during the welding process; By comparing the abnormal temperature points with the standard temperature curve of the welding process, the time periods in which the temperature deviated from the normal range can be identified. Based on the type and degree of temperature deviation, determine the specific process stage at which the defect may occur. The process stage includes the preheating stage, the welding stage, and the cooling stage. By combining information on the location of the weld joint and the stage at which the defect occurred, a complete result of defect localization and stage analysis is formed.

8. The method for inspecting the welding quality of flexible circuit boards for audio equipment according to claim 1, characterized in that, The process involves generating a welding quality assessment report based on the quality grade label and defect location information of each weld joint, including repair priorities and repair recommendations. Statistically analyze the quality grade distribution of all weld joints, and calculate the proportion of qualified weld joints and the number of various defective weld joints; Based on the defect type and severity, each defective solder joint is assigned a repair priority score, with a higher priority score indicating that it needs to be repaired earlier. For each type of defect, retrieve the corresponding repair process parameters and key operating points from the repair knowledge base; Based on the location information of specific solder joints, a personalized repair path plan is generated to avoid affecting other solder joints during the repair process; The statistical results of quality grade distribution, the priority ranking of defective weld repairs, personalized repair suggestions, and repair path planning are integrated to form a structured welding quality assessment report.

9. The method for inspecting the welding quality of flexible circuit boards for audio equipment according to claim 7, characterized in that, The process of comparing abnormal temperature points with the standard temperature curve of the welding process to identify time periods when the temperature deviates from the normal range includes: Retrieve the standard welding temperature curve corresponding to the current product model from the welding process parameter database. The standard welding temperature curve defines the ideal temperature range and duration for each process stage. Align the temperature distribution time-series curve of the weld point to be tested with the standard welding temperature curve on the time axis; Calculate the temperature difference between the two curves at each time point to form a temperature difference curve; Set a temperature difference threshold. When the temperature difference exceeds the threshold, mark that time point as a temperature anomaly. Analyze the distribution of continuous temperature anomalies and identify the duration of temperature anomalies; The specific stage at which the defect occurred is determined based on the duration of the temperature anomaly.

10. The method for inspecting the welding quality of flexible circuit boards for audio equipment according to claim 8, characterized in that, For each defect type, the process of retrieving corresponding repair process parameters and operational points from the repair knowledge base includes: Establish a mapping table between defect types and repair processes. The mapping table records the repair methods, required tools, temperature parameters, and time parameters corresponding to different defect types. Based on the defect type identified in the defect type probability distribution, the corresponding repair method is found in the mapping table; Obtain detailed operating steps, safety precautions, and solutions to common problems for this repair method from the repair knowledge base; Based on the specific location characteristics of the solder joint and the layout of surrounding components, adjust the local parameters in the repair process parameters; Generate a personalized repair operation guide that includes a repair step sequence, parameter settings, tool selection, and risk warnings.

Images