On-line judging system for steel mesh welding spot quality based on acoustic emission monitoring

Abstract

The application discloses a steel mesh welding spot quality online judgment system based on acoustic emission monitoring and relates to the technical field of welding spot quality detection.The application solves the technical problems of the precision and adaptability of feature extraction, the generalization ability and recognition precision of a judgment model, and the like, adopts a phased multi-domain feature extraction method, divides signals according to preheating-molten core-cooling stages, calculates energy features, extracts frequency features such as main frequency and spectral barycenter, removes weak correlation and redundant features through feature screening, redundancy removal and standardization processing, ensures the high information density and anti-interference ability of feature vectors, adapts to fluctuations of different processes and material batches, constructs a three-level fusion judgment architecture, fuses acoustic emission core features and multi-source complementary features such as vision and process parameters, an XGBoost model adapts to medium-complexity quality scenes, a lightweight CNN-LSTM model accurately identifies difficult-to-identify defects such as micro-cracks and micro-pores, and the recognition rate of defect types and the system judgment accuracy are improved.

Description

Technical Field

[0001] This invention relates to the field of weld joint quality inspection technology, specifically to an online weld joint quality assessment system for steel mesh based on acoustic emission monitoring. Background Technology

[0002] As a core connection structure for products such as automobiles, aerospace, and electronic equipment, the welding quality of steel mesh welds directly determines the structural strength, service life, and operational safety of the products. Therefore, accurate online assessment of the quality of steel mesh welds is a key link in zero-defect manufacturing in industrial production.

[0003] Currently, the mainstream online weld quality assessment technology in the industry is based on acoustic emission monitoring. Its principle is to capture the elastic wave signals generated during the weld nugget growth and defect initiation process, and combine the signal feature analysis to achieve quality assessment. However, the existing assessment technology has the following limitations:

[0004] The common practice of using a single-domain feature extraction method does not take into account the stage differences in welding sequence, and the feature calculation relies heavily on absolute values, without considering the impact of process fluctuations and material batch differences. This results in a blurred mapping relationship between features and weld quality, and weak anti-interference ability and adaptability.

[0005] In terms of judgment models and processes, most of them use a single threshold judgment or a single machine learning model, lacking hierarchical judgment logic; they do not integrate complementary information from multiple sources such as vision, process parameters, and thermal imaging, resulting in insufficient accuracy in identifying difficult-to-identify defects such as microcracks and micropores; at the same time, they lack dynamic calibration and weight optimization mechanisms, resulting in poor model generalization ability and easy misjudgment and missed detection due to changes in working conditions. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides an online quality assessment system for steel mesh weld points based on acoustic emission monitoring, which solves the problems of insufficient accuracy and adaptability of feature extraction, and defects in the generalization ability and recognition accuracy of the assessment model.

[0007] To achieve the above objectives, the present invention provides the following technical solution: an online quality assessment system for steel mesh weld points based on acoustic emission monitoring, comprising:

[0008] The adaptive processing module is used to acquire the preprocessed signal and extract the acoustic core features, including energy features and frequency features, from it. It then combines the pre-set rule base to complete the first-level judgment and outputs the first-level judgment information to the feature judgment and analysis module.

[0009] The feature determination and analysis module is used to perform secondary machine learning determination and tertiary deep learning determination on the primary determination information. For the secondary machine learning determination, it obtains the core features of vocalization and multi-source complementary features and removes abnormal features to generate an initial feature set. Based on the correlation coefficient, it filters collinear features to obtain a redundant feature set. It assigns weights to the core features of vocalization and multi-source complementary features and calculates the weighted sum to obtain a weighted feature vector. At the same time, it inputs the XGBoost model and calculates the confidence level to output the secondary determination result.

[0010] For the three-level deep learning judgment, the acoustic emission time-frequency map and visual image are obtained, the features of the solder joint area and surface texture are extracted, and the output is fed into the CNN-LSTM model. The output includes multiple predicted labels including quality label, defect type label and severity label. The three-level judgment result is generated by combining the confidence score.

[0011] The secondary and tertiary judgment results are then sent to the comprehensive judgment generation module.

[0012] The comprehensive judgment generation module is used to perform weighted fusion calculation on the first-level, second-level and third-level judgment results, and generate a comprehensive judgment result to the judgment information output module.

[0013] As a further aspect of the present invention, it also includes an acoustic emission signal acquisition module for capturing acoustic emission signals during the solder joint formation process and performing preprocessing, and outputting the preprocessed signal to the adaptive processing module.

[0014] The judgment information output module is used to display the comprehensive judgment results to the management personnel.

[0015] As a further aspect of the present invention, the energy feature is specifically extracted in the following way:

[0016] Obtain the preprocessed signal x(t) and the corresponding sampling frequency f. s And obtain the number of sampling points N, and use the discrete-time energy calculation formula E total = The total signal energy E was calculated. total Where i = 1, 2, ..., N represents the number of sampling points, and based on the stage boundary timestamps, the signal is segmented into a preheating segment x. pre (t), melting segment x nuc (t), Cooling section x cool (t), and calculate the energy of each stage respectively:

[0017] Among them, the preheating section x pre The energy corresponding to (t) is E. pre = , melting segment x nuc The energy corresponding to (t) is E. nuc = Cooling section xcool The energy corresponding to (t) is E. cool = .

[0018] As a further aspect of the present invention, the frequency feature is extracted in the following specific way:

[0019] Perform an FFT (Fourier Transform) on the preprocessed signal and calculate the power spectral density PSD(f) = Simultaneously, the frequency corresponding to the maximum value of PSD(f) is taken as the main frequency f0, and the frequency sequence f is obtained based on FFT. a and the corresponding power spectrum PSD(f a ), where a = 1, 2, ..., N / 2, according to the formula f c = The spectral centroid f is calculated. c .

[0020] As a further aspect of the present invention, the method for outputting the first-level determination information is as follows:

[0021] The obtained energy characteristics and frequency characteristics are compared with the preset rule library. The preset rule library stores the range of energy characteristics and frequency characteristics corresponding to different quality levels of solder joints. If all characteristics fall within the qualified range, it is judged as Level 1 qualified. If all characteristics fall within the unqualified range, it is judged as Level 1 pending review. If some characteristics fall within the qualified range, it is judged as Level 1 unqualified, and Level 1 judgment information is generated at the same time.

[0022] As a further aspect of the present invention, the method for outputting the secondary determination result is as follows:

[0023] Acquire the acoustic core features and multi-source complementary features corresponding to the preprocessed signal, where the multi-source complementary features include the splash area ratio S. r Current synchronization C s Cooling rate T v and the peak temperature of the melting nucleus T peak Remove constant features and abnormal features that exceed reasonable ranges, calculate the Pearson correlation coefficient R between each feature and the melt core diameter and tensile shear strength, retain features with |R|≥0.75, remove features with |R|<0.5, and output a 25-30 dimensional initial feature set;

[0024] Calculate the correlation coefficient matrix Corr of the initial feature set, identify highly collinear feature groups with a correlation coefficient ≥ 0.85, perform PCA dimensionality reduction on the collinear feature groups, extract one principal component with a variance contribution of ≥ 85% to replace the original feature group, and output an 18-22 dimensional feature set with redundancy removed.

[0025] As a further aspect of the present invention, weights are assigned to the core features and multi-source complementary features, and the weighted sum is used to obtain a weighted feature vector. The top 20 features with a cumulative weight ratio of ≥95% are retained, and a weighted feature vector F is generated. weighted According to formula F std =(F weighted - ) / Perform standardization calculations to obtain the standardized feature vector F. std ,in and The characteristic mean and standard deviation of qualified samples in the calibration library;

[0026] The obtained standardized feature vector F std The input is fed into the XGBoost model, which outputs the class probability and defect type probability. The class probability includes P. 合格 and P 不合格 The probability of defect type includes P 未熔合 P 气孔 P 裂纹 P 飞溅过量 According to the formula Score2=max(P) 合格 P 不合格） ×100 calculates the confidence level, where max(*) represents the maximum value function, specifically if P 合格 If the score is ≥0.95, the corresponding confidence level Score2 is ≥95 points, and the secondary judgment result is output.

[0027] As a further aspect of the present invention, the method for generating the three-level judgment result is as follows:

[0028] Acoustic emission time-frequency map and visual image are acquired. The core frequency region of the acoustic emission time-frequency map is retained and cropped, and Gaussian smoothing is performed. Then, the YOLOv8 lightweight model is used to perform target detection on the visual image, extracting the solder joint area and the corresponding surface texture features. The lightweight CNN-LSTM model is used as the standard. The spatial features of the acoustic emission time-frequency map and the surface features of the visual image are extracted by the CNN module, and the temporal evolution features of the acoustic emission signal are captured by the LSTM module. The model is combined to output multiple prediction labels, including quality label, defect type label and severity label.

[0029] Next, spatial feature similarity and temporal feature matching degree are calculated. Spatial feature similarity represents the cosine similarity between the features extracted by the model and the standard defect feature library. Temporal feature matching degree represents the normalized value of the dynamic time warp distance between the temporal sequence output by STM and the defect evolution template. The confidence score is calculated according to the formula Score3 = (spatial feature similarity + temporal feature matching degree) / 2 × 100. The three-level judgment result is output in combination with the confidence score.

[0030] As a further aspect of the present invention, the method for generating the comprehensive judgment result is as follows:

[0031] When the Level 1 assessment is deemed satisfactory, Score1 = 100 points; when it is deemed unsatisfactory, Score1 = 0 points; and when it is pending review, Score1 = 50 points. Weights W1, W2, and W3 are assigned to the Level 1, Level 2, and Level 3 assessment results, respectively, according to the formula Score1 = 100 points. total =Score1×W1+Score2×W2+Score3×W3, calculate the overall confidence score Score total ;

[0032] If Score total A score of ≥85 will generate a passing result. total If the score is ≤20, a non-compliant result will be generated; if the score is 20... <Score total If the score is less than 85, a manual review result will be generated, along with a comprehensive judgment result.

[0033] This invention provides an online quality assessment system for steel mesh weld points based on acoustic emission monitoring. Compared with existing technologies, it has the following advantages:

[0034] This invention employs a phased, multi-domain feature extraction method, segmenting the signal according to the preheating, melting, and cooling stages and calculating energy features. Combined with FFT transform, it accurately extracts frequency features such as the main frequency and spectral centroid. Through feature filtering, redundancy removal, and standardization, weakly correlated and redundant features are eliminated, ensuring high information density and anti-interference capability of the feature vectors. This adapts to the fluctuations in different processes and material batches. A three-level fusion judgment architecture is constructed, fusing acoustic emission core features with complementary features from multiple sources such as vision and process parameters. The XGBoost model is adapted to moderately complex quality scenarios, and the lightweight CNN-LSTM model accurately identifies difficult-to-identify defects such as microcracks and micropores, improving the defect type recognition rate and system judgment accuracy. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of the online steel mesh weld quality assessment system of the present invention.

[0036] Figure 2 This is a flowchart of the online quality assessment system for steel mesh weld points according to the present invention. Detailed Implementation

[0037] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0038] First Embodiment

[0039] Please see Figure 1 and Figure 2 This application provides an online quality assessment system for steel mesh weld points based on acoustic emission monitoring, including an acoustic emission signal acquisition module, an adaptive processing module, a feature judgment and analysis module, a comprehensive evaluation generation module, and a judgment information output module, and combines... Figure 1 It can be seen that the information between the above functional modules is transmitted in one direction.

[0040] The acoustic emission signal acquisition module is used to capture acoustic emission signals during the solder joint formation process using high-sensitivity piezoelectric sensors. It employs four sets of high-sensitivity piezoelectric sensors, with each monitoring unit consisting of 5-8 solder joints. The sensors are arranged in an equilateral triangle and rigidly fixed to the workpiece using a high-temperature resistant coupling agent. The module amplifies, filters, and reduces noise in the acquired acoustic emission signals to obtain a pre-processed signal. The specific processing method is as follows:

[0041] First, a differential preamplifier is connected to suppress common-mode interference through the differential amplification structure. At the same time, power frequency notch filter and bandpass filter circuits are integrated to specifically filter out low-frequency noise such as industrial power grid interference and mechanical vibration, as well as high-frequency noise generated by sensor resonance, thus initially improving the signal-to-noise ratio. Then, the signal is decomposed into 4-6 modal components through VMD. Based on the energy proportion and frequency distribution of each mode, the dominant mechanical vibration mode <200Hz is automatically eliminated. Then, the Birgé-Massart adaptive threshold strategy is adopted to dynamically adjust the threshold based on the local signal-to-noise ratio of the signal. While preserving the defect characteristics, random noise is suppressed to the maximum extent. Finally, a high-fidelity preprocessed signal is obtained and transmitted to the adaptive processing module.

[0042] An adaptive processing module is used to extract the corresponding acoustic core features based on the obtained preprocessed signal. The acoustic core features include energy features and frequency features, and the specific extraction method is as follows:

[0043] The energy feature extraction method involves acquiring the preprocessed signal x(t) and the corresponding sampling frequency f. s And obtain the number of sampling points N, and use the discrete-time energy calculation formula E total = The total signal energy E was calculated. totalWhere i = 1, 2, ..., N represents the number of sampling points, and based on the stage boundary timestamps, the signal is segmented into a preheating segment x. pre (t), melting segment x nuc (t), Cooling section x cool (t), and calculate the energy of each stage respectively:

[0044] Among them, the preheating section x pre The energy corresponding to (t) is E. pre = , melting segment x nuc The energy corresponding to (t) is E. nuc = Cooling section x cool The energy corresponding to (t) is E. cool = ;

[0045] The frequency feature extraction method involves performing an FFT (Fourier Transform) on the preprocessed signal and calculating the power spectral density PSD(f) = Simultaneously, the frequency corresponding to the maximum value of PSD(f) is taken as the main frequency f0, and the frequency sequence f is obtained based on FFT. a and the corresponding power spectrum PSD(f a ), where a = 1, 2, ..., N / 2, according to the formula f c = The spectral centroid f is calculated. c Simultaneously, the obtained energy characteristics and frequency characteristics are compared and judged with the preset rule base. The preset rule base stores the range of energy characteristics and frequency characteristics corresponding to different quality levels of solder joints. If all characteristics fall within the qualified range, it is judged as Level 1 qualified. If all characteristics fall within the unqualified range, it is judged as Level 1 pending review. If some characteristics fall within the qualified range, it is judged as Level 1 unqualified. At the same time, Level 1 judgment information is generated and transmitted to the feature judgment and analysis module.

[0046] Second Embodiment

[0047] As a second embodiment of the present invention, it is implemented based on the first embodiment, and the difference from the first embodiment is as follows:

[0048] The feature determination and analysis module is used to perform secondary and tertiary determination and analysis on the acquired primary determination information. The secondary determination and analysis process is as follows:

[0049] First, the acoustic core features and multi-source complementary features corresponding to the preprocessed signal are obtained. Among them, the multi-source complementary features include the splash area ratio S. r Current synchronization C s Cooling rate T vand the peak temperature of the melting nucleus T peak Next, constant value characteristics caused by sensor malfunctions and abnormal characteristics exceeding reasonable ranges are eliminated, such as energy = 0 or splash area percentage S. r >50%, and calculate the Pearson correlation coefficient R between each feature and the quality index, where the quality index includes the melt core diameter and tensile shear strength. At the same time, retain features with R≥0.75, remove weakly correlated features with R<0.5, and output the initial feature set with a dimension of 25-30.

[0050] The correlation coefficient matrix Coor of the initial feature set is calculated, and highly collinear feature groups with correlation coefficients ≥ 0.85 are identified. PCA is then performed on the obtained collinear feature groups to extract a principal component with a variance contribution ≥ 85%, which replaces the original feature group, outputting a redundant feature set with dimensions 18-22. Next, the weights of the core acoustic features and multi-source complementary features are determined, and the features are summed with their corresponding weights to obtain a weighted feature vector. The top 20 features with a cumulative weight percentage ≥ 95% are retained, and a weighted feature vector F is generated. weighted According to formula F std =(F weighted - ) / Perform standardization calculations to obtain the standardized feature vector F. std ,in and The characteristic mean and standard deviation of qualified samples in the calibration library;

[0051] Load the XGBoost model with the following parameters: learning rate 0.1, tree depth 6, number of iterations 200, minimum weight of leaf nodes 1, and output the normalized feature vector F. std The input is fed into the XGBoost model, which outputs two types of results: class probability and defect type probability. The class probability includes P... 合格 and P 不合格 The probability of defect type includes P 未熔合 P 气孔 P 裂纹 P 飞溅过量 Then, according to the formula Score2=max(P) 合格 P 不合格） ×100 calculates the confidence level, where max(*) represents the maximum value function, specifically if P 合格 If the score is ≥0.95, the corresponding confidence level Score2 is ≥95 points. At the same time, the secondary judgment result is output, which includes: judgment label: qualified / unqualified, defect type: none / not fused, etc., confidence level: Score2;

[0052] The three-level judgment analysis and processing method is as follows: Acoustic emission time-frequency map and visual image are obtained. The core frequency region of the acoustic emission time-frequency map is retained and cropped, and Gaussian smoothing is performed. Then, the YOLOv8 lightweight model is used to perform target detection on the visual image, extracting the solder joint area and the corresponding surface texture features. The lightweight CNN-LSTM model is used as the standard. The structure of the CNN-LSTM model includes an input layer, CNN1, ReLU, pooling layer 1, CNN2, ReLU, pooling layer 2, LSTM1, LSTM2, fully connected layer, and output layer. The spatial features of the acoustic emission time-frequency map and the surface features of the visual image are extracted through the CNN module. The temporal evolution features of the acoustic emission signal are captured through the LSTM module. The model is combined to perform calculations and output multiple prediction labels, including quality labels, defect type labels, and severity labels. The severity labels include mild, moderate, and severe.

[0053] Next, spatial feature similarity and temporal feature matching degree are calculated. Spatial feature similarity represents the cosine similarity between the features extracted by the model and the standard defect feature library. Temporal feature matching degree represents the normalized value of the dynamic time warp distance between the temporal sequence output by STM and the defect evolution template. The confidence score is calculated according to the formula Score3 = (spatial feature similarity + temporal feature matching degree) / 2 × 100. If similarity + matching degree ≥ 1.7, Score3 ≥ 85 points; if ≤ 1.0, Score3 ≤ 50 points. At the same time, the three-level judgment result is output in combination with the confidence score, specifically including judgment label: qualified / unqualified, defect type: none / not fused, etc., severity: none / mild, etc., and confidence score: Score3.

[0054] At the same time, the generated secondary and tertiary judgment results are transmitted to the comprehensive judgment generation module.

[0055] Third Embodiment

[0056] As a third embodiment of the present invention, it is implemented based on the second embodiment, and the difference from the second embodiment is as follows:

[0057] The comprehensive judgment generation module performs a comprehensive quality judgment based on the results of Level 1, Level 2, and Level 3 judgments. It assigns weights to each level's judgment result and calculates the confidence level corresponding to the Level 1 judgment result: if the Level 1 judgment is satisfactory, Score1 = 100 points; if it is unsatisfactory, Score1 = 0 points; if it is pending review, Score1 = 50 points. The module then performs a weighted summation based on the confidence levels of Level 2 and Level 3 judgments, according to the formula Score1. total =Score1×W1+Score2×W2+Score3×W3, calculate the overall confidence score Score total, where W1, W2 and W3 are the corresponding weights;

[0058] The overall confidence score is calculated based on the score. total A comprehensive result judgment is made, if the Score total A score of ≥85 will generate a passing result. total If the score is ≤20, a non-compliant result will be generated; if the score is 20... <Score total If the score is less than 85, a manual review result will be generated, and the generated comprehensive judgment result will be transmitted to the judgment information output module.

[0059] The judgment information output module is used to display the acquired judgment information to the corresponding management personnel.

[0060] Fourth embodiment

[0061] As a fourth embodiment of the present invention, the focus is on combining the implementation processes of the first, second and third embodiments.

[0062] Some of the data in the above formulas are numerical calculations with dimensions removed, and the contents not described in detail in this specification are all prior art known to those skilled in the art.

[0063] 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. An online quality assessment system for steel mesh weld points based on acoustic emission monitoring, characterized in that, include: The adaptive processing module is used to acquire the preprocessed signal and extract the acoustic core features, including energy features and frequency features, from it. It then combines the pre-set rule base to complete the first-level judgment and outputs the first-level judgment information to the feature judgment and analysis module. The feature determination and analysis module is used to perform secondary machine learning determination and tertiary deep learning determination on the primary determination information. For the secondary machine learning determination, it obtains the core features of vocalization and multi-source complementary features and removes abnormal features to generate an initial feature set. Based on the correlation coefficient, it filters collinear features to obtain a redundant feature set. It assigns weights to the core features of vocalization and multi-source complementary features and calculates the weighted sum to obtain a weighted feature vector. At the same time, it inputs the XGBoost model and calculates the confidence level to output the secondary determination result. For the three-level deep learning judgment, the acoustic emission time-frequency map and visual image are obtained, the features of the solder joint area and surface texture are extracted, and the output is fed into the CNN-LSTM model. The output includes multiple predicted labels including quality label, defect type label and severity label. The three-level judgment result is generated by combining the confidence score. The secondary and tertiary judgment results are then transmitted to the comprehensive judgment generation module. The comprehensive judgment generation module is used to perform weighted fusion calculation on the first-level, second-level and third-level judgment results, generate a comprehensive judgment result and transmit it to the judgment information output module.

2. The online steel mesh weld quality assessment system based on acoustic emission monitoring according to claim 1, characterized in that, It also includes an acoustic emission signal acquisition module for capturing and preprocessing acoustic emission signals during the solder joint formation process, and outputting the preprocessed signal to the adaptive processing module; The judgment information output module is used to display the comprehensive judgment results to the management personnel.

3. The online quality assessment system for steel mesh weld points based on acoustic emission monitoring according to claim 1, characterized in that, The specific methods for extracting energy features are as follows: Obtain the preprocessed signal x(t) and the corresponding sampling frequency f. s And obtain the number of sampling points N, and use the discrete-time energy calculation formula E total = The total signal energy E was calculated. total Where i = 1, 2, ..., N represents the number of sampling points, and based on the stage boundary timestamps, the signal is segmented into a preheating segment x. pre (t), melting segment x nuc (t), Cooling section x cool (t), and calculate the energy of each stage respectively: Among them, the preheating section x pre The energy corresponding to (t) is E. pre = , melting segment x nuc The energy corresponding to (t) is E. nuc = Cooling section x cool The energy corresponding to (t) is E. cool = .

4. The online steel mesh weld quality assessment system based on acoustic emission monitoring according to claim 1, characterized in that, The specific methods for extracting frequency features are as follows: Perform an FFT transform on the preprocessed signal and calculate the power spectral density PSD(f) = Simultaneously, the frequency corresponding to the maximum value of PSD(f) is taken as the main frequency f0, and the frequency sequence f is obtained based on FFT. a and the corresponding power spectrum PSD(f a ), where a = 1, 2, ..., N / 2, according to the formula f c = The spectral centroid f is calculated. c .

5. The online steel mesh weld quality assessment system based on acoustic emission monitoring according to claim 1, characterized in that, The method for outputting the first-level judgment information is as follows: The obtained energy characteristics and frequency characteristics are compared with the preset rule library. The preset rule library stores the range of energy characteristics and frequency characteristics corresponding to different quality levels of solder joints. If all characteristics fall within the qualified range, it is judged as Level 1 qualified. If all characteristics fall within the unqualified range, it is judged as Level 1 pending review. If some characteristics fall within the qualified range, it is judged as Level 1 unqualified, and Level 1 judgment information is generated at the same time.

6. The online steel mesh weld quality assessment system based on acoustic emission monitoring according to claim 1, characterized in that, The method for outputting the secondary judgment result is as follows: Acquire the acoustic core features and multi-source complementary features corresponding to the preprocessed signal, where the multi-source complementary features include the splash area ratio S. r Current synchronization C s Cooling rate T v and the peak temperature of the melting nucleus T peak Remove constant features and abnormal features that exceed reasonable ranges, calculate the Pearson correlation coefficient R between each feature and the melt core diameter and tensile shear strength, retain features with |R|≥0.75, remove features with |R|<0.5, and output a 25-30 dimensional initial feature set; Calculate the correlation coefficient matrix Corr of the initial feature set, identify highly collinear feature groups with a correlation coefficient ≥ 0.85, perform PCA dimensionality reduction on the collinear feature groups, extract one principal component with a variance contribution of ≥ 85% to replace the original feature group, and output an 18-22 dimensional feature set with redundancy removed.

7. The online steel mesh weld quality assessment system based on acoustic emission monitoring according to claim 6, characterized in that, Weights are assigned to the core features and multi-source complementary features, and the weighted sum is used to obtain a weighted feature vector. The top 20 features with a cumulative weight ratio of ≥95% are retained, and a weighted feature vector F is generated. weighted According to formula F std =(F weighted - ) / Perform standardization calculations to obtain the standardized feature vector F. std ,in The characteristic mean and standard deviation of qualified samples in the calibration library; The obtained standardized feature vector F std The input is fed into the XGBoost model, which outputs the class probability and defect type probability. The class probability includes P. 合格 and P 不合格 The probability of defect type includes P 未熔合 P 气孔 P 裂纹 P 飞溅过量 According to the formula Score2=max(P) 合格 P 不合格 The confidence level is calculated by multiplying P by 100, where max(*) represents the maximum value function. Specifically, if P... 合格 If the score is ≥0.95, the corresponding confidence level Score2 is ≥95 points, and the secondary judgment result is output.

8. The online steel mesh weld quality assessment system based on acoustic emission monitoring according to claim 1, characterized in that, The method for generating the three-level judgment result is as follows: Acoustic emission time-frequency map and visual image are acquired. The core frequency region of the acoustic emission time-frequency map is retained and cropped, and Gaussian smoothing is performed. Then, the YOLOv8 lightweight model is used to perform target detection on the visual image, extracting the solder joint area and the corresponding surface texture features. The lightweight CNN-LSTM model is used as the standard. The spatial features of the acoustic emission time-frequency map and the surface features of the visual image are extracted by the CNN module, and the temporal evolution features of the acoustic emission signal are captured by the LSTM module. The model is combined to output multiple prediction labels, including quality label, defect type label and severity label. Next, spatial feature similarity and temporal feature matching degree are calculated. Spatial feature similarity represents the cosine similarity between the features extracted by the model and the standard defect feature library. Temporal feature matching degree represents the normalized value of the dynamic time warp distance between the temporal sequence output by STM and the defect evolution template. The confidence score is calculated according to the formula Score3 = (spatial feature similarity + temporal feature matching degree) / 2 × 100. The three-level judgment result is output in combination with the confidence score.

9. The online steel mesh weld quality assessment system based on acoustic emission monitoring according to claim 1, characterized in that, The method for generating the comprehensive judgment result is as follows: When the Level 1 assessment is deemed satisfactory, Score1 = 100 points; when it is deemed unsatisfactory, Score1 = 0 points; and when it is pending review, Score1 = 50 points. Weights W1, W2, and W3 are assigned to the Level 1, Level 2, and Level 3 assessment results, respectively, according to the formula Score1 = 100 points. total =Score1×W1+Score2×W2+Score3×W3, calculate the overall confidence score Score total ; If Score total A score of ≥85 will generate a passing result. total If the score is ≤20, a non-compliant result will be generated; if the score is 20... <Score total If the score is less than 85, a manual review result will be generated, along with a comprehensive judgment result.

Images