Busbar laser welding pseudo soldering defect online diagnosis method and system based on photoelectric signals

Abstract

The invention belongs to but is not limited to the technical field of defect diagnosis, particularly relates to an on-line diagnosis method and system for a busbar laser welding pseudo soldering defect based on a photoelectric signal, and provides a feature definition method based on a reference value. According to the feature definition method, the diagnosis accuracy is improved through multi-band data segmentation and calculation of the signal difference coefficient feature matrix, and the method has the advantages of adapting to different process conditions and checking local defects. On the basis, a quality diagnosis model based on the random forest is constructed. By means of the method, whether the cold joint defect exists in busbar laser welding or not can be detected, induction factors of the cold joint defect can be further diagnosed, and effective support is provided for rapidly positioning problem sources and taking targeted measures in industrial production.

Description

Technical Field

[0001] The present invention belongs to but is not limited to the field of defect diagnosis technology, and in particular relates to an online diagnosis method and system for bus laser welding cold weld defects based on photoelectric signals. Background Art

[0002] As an important conductive device in the power system, busbars are widely used in distribution networks, industrial automation, electric vehicles and other fields. Busbars are usually welded to achieve efficient electrical connections. However, busbars are often used to carry large-scale current transmission, which means a large number of welding processes. For example, in automotive power batteries, a battery module contains thousands of cells that need to be connected to each other, which poses a challenge to the efficiency of welding and the high reliability and consistency of welding quality.

[0003] Laser welding is widely used in the electrical connection of busbars due to its advantages of high precision, high speed and low heat-affected zone, and can achieve low contact resistance and high joint strength. However, during the processing process, factors such as insufficient laser power caused by equipment aging and failure, changes in the overlap gap caused by loose fixtures, unstable distance between the welding head and the workpiece caused by repeated positioning errors, and surface contamination of the workpiece caused by processing residues and poor storage environment may lead to laser welding quality problems, especially cold welding defects. Cold welding defects not only increase contact resistance and reduce power supply capacity, but may also cause welds to fall off, seriously affecting the performance and safety of the busbar. Therefore, it is crucial to detect cold welding defects in a timely manner. Through automated quality inspection, defects can be discovered and corrected in a timely manner, rework and scrap rates can be reduced, production efficiency can be improved, and costs can be reduced.

[0004] In recent years, advanced sensing technologies have been used for busbar quality diagnosis, such as spectrometers, high-speed cameras, optical coherence tomography, photodiodes, etc., which have the advantages of non-contact and high real-time performance. These sensing methods are often combined with image processing, statistical methods, machine learning / deep learning and other technologies to provide solutions for laser welding process signal monitoring and defect diagnosis.

[0005] However, the camera is easily affected by smoke, dust and light changes in the environment, which affects the image quality and increases the challenge of real-time processing. Spectral data analysis is complex, the spectrometer has a limited monitoring area, and is easily affected by noise, external light and dynamic changes in welding, resulting in reduced detection accuracy. The limitations of OCT for welding quality diagnosis include high equipment cost, complex operation, the need for precise re-alignment when process parameters change, and sensitivity to vibration and temperature changes. Welding quality inspection and diagnosis technology using photodiode sensors has the advantages of low equipment cost, flexible configuration, simple structure, high sampling frequency, and fast processing speed. It has application advantages in quality monitoring of the actual industrial production process of the bus.

[0006] The existing laser welding quality diagnosis technology based on photoelectric signals currently mostly uses the threshold judgment method in industrial practice, that is, the welding quality is judged by manually setting the threshold to quantitatively describe the extent to which the signal curve generated by the measured workpiece exceeds the envelope calculated by the good weld. The accuracy of such a method depends on the human factor of the threshold setting, and the reliability is poor and the intelligence is relatively low. In recent years, some researchers have tried to use machine learning algorithms combined with photoelectric signal characteristics to perform intelligent diagnosis of welding quality. However, these studies usually extract the characteristics of the photoelectric signal itself, such as extracting statistical characteristics such as the mean and divergence of signals in different bands, or using wavelet packet transform analysis and probability density analysis to extract the characteristic values ​​of the original photoelectric signal, reflecting the corresponding laser welding quality status. The characteristic of these methods is that they only extract features from the photoelectric signal itself generated during each laser welding process. Since the photoelectric signal does not directly measure the size of the weld, the interaction of process parameters will interfere with the results of quality diagnosis. In addition, the light radiation characteristics generated by laser welding under different process conditions vary greatly. For example, the laser power will change the intensity of the light radiation signal, and the difference in the overlapping materials will affect the characteristics of the back reflection. These factors make the photoelectric signal characteristic interval judged as a good weld under different process conditions may be different. If a large amount of data needs to be re-acquired to train the quality assessment model under each process condition, this will not meet the requirements of high efficiency and high reliability for busbar laser welding quality diagnosis.

[0007] In view of the above analysis, the technical problems that urgently need to be solved in the existing technology are: the accuracy of the existing threshold method depends on the human factor of threshold setting, with poor reliability and low intelligence; the feature input of the existing machine learning method depends on the signal itself, and the accuracy and adaptability to different process conditions are insufficient. Summary of the invention

[0008] In view of the defects and shortcomings in the existing laser welding quality diagnosis technology, the present invention proposes an online diagnosis method and system for bus laser welding cold weld defects based on photoelectric signals. In the proposed method, a feature definition method based on reference values ​​is proposed. This feature definition method improves the accuracy of diagnosis through multi-band data segmentation and calculation of signal difference coefficient feature matrix, and has the advantages of adapting to different process conditions and checking local defects. On this basis, a quality diagnosis model based on random forest is constructed. This method can not only detect whether there is a cold weld defect in bus laser welding, but also further diagnose the inducing factors of the cold weld defect, which provides effective support for quickly locating the source of the problem and taking targeted measures in industrial production.

[0009] The present invention is implemented as follows: an online diagnosis method for busbar laser welding cold welding defects based on photoelectric signals, comprising:

[0010] S1, raw photoelectric signal data processing;

[0011] S2, feature definition based on reference values;

[0012] S3, build a quality diagnosis model based on random forest.

[0013] Furthermore, S1 specifically includes: during each laser welding process, the photodiode records a sequence of light signal sampling points in three bands. Each sampling point contains time, band and its corresponding signal strength information. In order to extract valid data, the invalid part where no welding occurs is first removed by setting a suitable threshold, and the time point when welding starts is calculated, thereby determining the valid interval of the signal.

[0014] In order to eliminate the high-frequency noise interference in the original signal, a low-pass filter is used to process the signal. Low-pass filtering can remove the high-frequency components in the signal while retaining the original characteristics of the low frequency, thereby smoothing the signal curve. The function of the low-pass filter is to filter the signal through formula 1.

[0015]

[0016] Among them, x(t) is the original signal, h(t) is the impulse response function of the filter, and common low-pass filters include ideal low-pass filter, Butterworth low-pass filter, etc.

[0017]

[0018] The frequency response of an ideal low-pass filter is shown in Equation 2. Where fc is the cutoff frequency of the filter, indicating the highest frequency passed. After filtering, a three-band "time-signal intensity" graph is plotted for further analysis.

[0019] Further, S2 specifically includes: calculating features by taking the quality assessment template under specific process conditions as a reference, thereby improving the reliability of the quality diagnosis model and enabling it to adapt to the welding process diagnosis under different process conditions. In order to fully reflect the laser welding process and avoid missing local defects, the present invention adopts a multi-band data segmentation method and calculates the signal difference coefficient of each band. Among them, the signal difference coefficient represents the difference ratio between the signal and the reference in each data segment, reflecting information such as the degree and direction of the difference between the signal and the reference. The signal difference coefficients are arranged in a signal difference coefficient-feature matrix according to the specified positions as the features of the input model, which not only increases the interpretability of the model, but also provides support for further subdivision of defects.

[0020] Furthermore, the specific calculation process of feature extraction is as follows.

[0021] The reference curve is XY, the measured sample curve is xy, and the total number of sample points is n.

[0022] First, the quality assessment template is calculated. Each sample corresponds to a good weld test sample group containing at least 15 samples under its process conditions. The average value of the intensity of the photoelectric signal data of all samples in the good weld test sample group within the same band range and at the same time point is calculated to generate a reference for the three bands of the sample.

[0023] Secondly, perform multi-band data segmentation. First, the photoelectric signals of the reference curve and the measured sample curve need to be aligned to ensure that they occur in the same time series, and then the sample data is segmented. Specifically, the photoelectric signal data is divided into 50 segments by time, and the basic sampling points a and remainder b of each segment are calculated to satisfy Formula 3.

[0024] n=50a+b, 0≤b<50 (3)

[0025] According to this segmentation rule, each of the first 49 segments contains a sampling points, and the number of sampling points in the 50th segment is a + b. If n is divisible by 50 (ie, b = 0), then all 50 segments contain a sampling points.

[0026] Next, calculate the signal difference coefficient in each data segment. Take the signal difference coefficient of the i-th data segment in the first band as an example, its position in the matrix is ​​(1, i). The calculation rule of the signal difference coefficient is as follows: calculate the ratio of the vertical coordinate difference of all sampling points between the measured sample curve and the reference curve in the data segment to the vertical coordinate of the corresponding sampling point of the reference curve, and take the average of these ratios. Assuming that the time coordinate value of the sampling point contained in this segment of data is from T1 to T2, the calculation of the signal difference coefficient satisfies Formula 4.

[0027]

[0028] Finally, the complete RDC-FM is obtained. The signal difference coefficients of all data segments in the three bands of the experimental sample are calculated to generate a 3x50 signal difference coefficient feature matrix. The signal difference coefficient feature matrix describes the degree of difference between the measured sample curve and the reference curve, covering information such as the amplitude, direction, time and band of the difference, and unifies the judgment criteria for good welds under different welding conditions. This method ensures that the entire welding process, including local cold weld defects, is captured in an accurate and adaptable manner to different welding conditions, providing important basic data for subsequent machine learning model training and welding quality diagnosis.

[0029] Furthermore, S3 specifically includes:

[0030] Dataset construction: Multiple sub-datasets are generated from the original training set through Bootstrap sampling. Each sub-dataset consists of some samples in the original dataset, and some samples may be repeated due to sampling with replacement. Assume that the training set D = {x_1, x_2, ..., x_N}, where xi represents the training sample and N is the number of samples. In the Bootstrap sampling process, N samples are randomly selected from the training set D each time, and repeated samples may appear.

[0031] Decision tree training: For each sub-dataset D1 generated by Bootstrap sampling, a decision tree is constructed. At each node of the tree, a subset of features is randomly selected (random selection of features) to find the optimal splitting feature. The node is recursively split until the stopping criterion (such as the depth of the tree or the number of samples in the leaf node) is met. The construction of the decision tree relies on Information Gain (IG) and Gini Index to measure the quality of the split.

[0032] Information gain is a measure of the improvement in class purity after splitting a data set by a certain feature. Assuming that feature A is selected for splitting at node n, the information gain calculation formula is:

[0033]

[0034] Among them, H(D) is the entropy of the data set D, which represents the uncertainty of the data set:

[0035]

[0036] Among them, p i is the probability of category i in the data set, m is the total number of categories; D v It is the data subset when feature A takes the value v.

[0037] The Gini index is used to measure the impurity of a data set. The smaller the Gini index, the higher the purity of the data set. The calculation formula of the Gini index is:

[0038]

[0039] Where pi is the probability of belonging to category i in dataset D.

[0040] Integration and voting: After each decision tree is trained, the prediction results of all decision trees are integrated by majority voting. In the test phase, for the input sample x, each decision tree Ti will output a category prediction Ci. The final prediction category C is determined by voting:

[0041] C=mode(C1,C2,...,C N) (8)

[0042] Among them, N is the number of decision trees in the random forest, and mode means selecting the category with the highest frequency.

[0043] Output: As the classification results of all decision trees are integrated, the final classification result is the output of the random forest. This integration method can significantly improve the classification performance and reduce the bias and variance that a single decision tree may produce on the training set.

[0044] Among them, N is the number of decision trees in the random forest, and mode means selecting the category with the highest frequency.

[0045] Furthermore, we tuned the two hyperparameters, the number of decision trees and the minimum number of leaves, using a grid search combined with Bayesian optimization.

[0046] First, we systematically explore the hyperparameter space through grid search to provide initial results for subsequent Bayesian optimization. The hyperparameter ranges of the grid search are set as follows: minimum number of leaves (1, 3, 5, 7, 9) and number of decision trees (50, 100, 150, 200, 250). Performance evaluation is performed through 5-fold cross validation on the data set. The optimal hyperparameter combination obtained by grid search is 1 minimum number of leaves and 100 decision trees.

[0047] On this basis, in order to further optimize the model performance, Bayesian optimization is used to tune within a finer range of hyperparameters. Specifically, the minimum number of leaves ranges from 1 to 6, and the number of decision trees ranges from 50 to 150. Bayesian optimization iteratively adjusts parameters based on the results of the previous round and eventually converges to the global optimal hyperparameter combination. The best observed parameter combination is a minimum number of leaves of 1 and a number of decision trees of 107. The best parameter combination obtained based on the surrogate model is a minimum number of leaves of 1 and a number of decision trees of 143. The final hyperparameter combination significantly improves the accuracy and robustness of the model, providing reliable support for diagnostic tasks.

[0048] Another object of the present invention is to provide an online diagnosis system for busbar laser welding cold weld defects based on photoelectric signals, which realizes the online diagnosis method for busbar laser welding cold weld defects based on photoelectric signals, comprising:

[0049] A data processing module, used for processing raw photoelectric signal data;

[0050] A feature definition module, used for defining features based on reference values;

[0051] The quality diagnosis model building module is used to build a quality diagnosis model based on random forest.

[0052] Another object of the present invention is to provide a computer device, which includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the online diagnosis method for bus laser welding cold soldering defects based on photoelectric signals.

[0053] Another object of the present invention is to provide a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the processor executes the steps of the online diagnosis method for bus laser welding cold weld defects based on photoelectric signals.

[0054] Another object of the present invention is to provide an information data processing terminal, which includes the above-mentioned bus laser welding cold weld defect online diagnosis system based on photoelectric signals.

[0055] In combination with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solutions to be protected by the present invention are as follows:

[0056] First, in view of the defects and shortcomings in the existing laser welding quality diagnosis technology, the present invention proposes an online diagnosis method for bus laser welding cold weld defects based on photoelectric signals. In the proposed method, a feature definition method based on reference values ​​is proposed. This feature definition method improves the accuracy of diagnosis through multi-band data segmentation and calculation of the signal difference coefficient feature matrix, and has the advantages of adapting to different process conditions and checking local defects. On this basis, a quality diagnosis model based on random forest is constructed. This method can not only detect whether there is a cold weld defect in bus laser welding, but also further diagnose the inducing factors of the cold weld defect, which provides effective support for quickly locating the source of the problem in industrial production and taking targeted measures.

[0057] Second, the online diagnosis method for busbar laser welding cold solder joint defects based on photoelectric signal driven machine learning proposed in the present invention uses photodiodes as sensors, which have significant advantages such as low cost, flexible configuration, and rapid response. Combined with these technical features, the present invention not only has strong technical advantages, but also can bring significant economic benefits and commercial value to multiple industries.

[0058] (1) Improve production efficiency and quality

[0059] This technology can detect cold soldering defects in the welding process in real time and diagnose the factors causing the defects, which significantly improves the automation level of the welding process and the reliability of product quality. The use of photodiode sensors makes the detection system not only low-cost and fast-response, but also flexible in configuration to adapt to different production environments and process requirements. These features enable this technology to significantly improve production efficiency, reduce defective product rates, reduce rework and scrap in large-scale manufacturing processes, thereby reducing overall production costs.

[0060] (2) Reduce operating costs

[0061] As a sensor, photodiodes have extremely high sensitivity and stability. They can quickly capture the signal of cold soldering defects during the welding process, thereby improving the accuracy and timeliness of defect detection. This advantage not only helps companies save the cost of repairs and replacements caused by defects, but also reduces equipment downtime and improves production continuity. In addition, with the improvement of welding quality and defect detection accuracy, the overall quality of the product will be guaranteed, customer satisfaction will also be improved, and the market competitiveness of the company will be further promoted.

[0062] (3) Promote industrial upgrading and technological innovation

[0063] With the development of electric vehicles, smart grids and other high-tech fields, the demand for efficient and reliable electrical connections is increasing. The online cold welding detection system of the present invention can not only meet the existing manufacturing needs, but also promote technological innovation in these industries in intelligent manufacturing and automated detection. Especially in the manufacturing process of electric vehicles, welding quality is crucial to the safety and performance of battery packs. The high response speed and flexible configuration of photodiode sensors enable this technology to achieve efficient and accurate defect diagnosis in different welding environments, further promoting the technological upgrading of related industries.

[0064] (4) Market Competitive Advantages and Demand Growth

[0065] The global society is facing challenges such as climate change and energy shortages, and the demand for electric vehicles is growing rapidly. In order to reduce greenhouse gas emissions, electric vehicles have become a global trend to replace traditional fossil fuel vehicles. According to estimates, by 2030, there will be at least 30 million zero-emission electric vehicles (EVs) on the road in the European Union alone. Assembling electric vehicle battery packs requires about 20,000 battery-to-tag welds, which means billions of welds will be performed every year. Improving welding quality and efficiency is crucial for electric vehicle manufacturers.

[0066] The online detection method for cold welding defects of the present invention has a wide range of application prospects, especially in the battery pack welding process. With the scale of electric vehicle production and the complexity of welding technology, the requirements for welding quality control are increasing day by day, and the present invention will become one of the indispensable key technologies in the electric vehicle industry.

[0067] (5) Investment potential and returns

[0068] The present invention can not only provide an efficient and low-cost welding quality inspection solution for existing manufacturing companies, but also has wide industry applicability. With the explosion of demand for electric vehicles and global attention to high-precision manufacturing technology, the commercialization of this technology will attract a large number of investors. It is expected that with the gradual promotion of technology and the expansion of market applications, the return on investment will be apparent in the short term, and sustained growth in profits can be achieved in the long term. Especially in high-growth fields such as electric vehicles and power systems, the return on investment of this technology is very attractive.

[0069] The technical solution of the present invention has significant advantages in improving welding quality, reducing production costs and improving efficiency by combining the advantages of low cost, high response and flexible configuration of photodiode sensors with the machine learning method of cold solder joint defect detection driven by photoelectric signals. With the rapid growth of demand for electric vehicles and global energy transformation, the present invention will bring important technological breakthroughs and business opportunities to the manufacturing industry, especially the electric vehicle industry. Its broad market prospects and technological innovation value will provide strong competitiveness for related companies and bring considerable returns to investors.

[0070] Third, the existing laser welding quality diagnosis technology based on photoelectric signals mostly relies on threshold judgment methods in practical industrial applications. This method manually sets a threshold to quantitatively describe the extent to which the signal curve generated by the workpiece exceeds the envelope of a good weld, and then judges the welding quality. However, the accuracy of this method is highly dependent on the setting of the threshold, and the setting of the threshold is usually affected by human factors, which makes it less reliable and less intelligent. With the industrial demand for higher intelligence and accuracy, the existing diagnostic methods have been unable to meet the needs of efficient welding quality diagnosis under complex process conditions.

[0071] In recent years, some researchers have tried to combine machine learning algorithms and photoelectric signal characteristics for intelligent diagnosis of welding quality. Although such methods have improved the accuracy of welding quality detection to a certain extent, most studies only extract the characteristics of the photoelectric signal itself, such as the mean and variance of signals in different bands, or use methods such as wavelet packet transform and probability density analysis to extract the characteristic values ​​of the photoelectric signal to reflect the state of laser welding quality. However, these methods have the following shortcomings:

[0072] (1) Process parameter interference: The photoelectric signal itself does not directly measure the size of the weld. Therefore, various process parameters in the welding process, such as laser power, focal position, and the characteristics of the welding material, will affect the photoelectric signal and thus affect the accuracy of quality diagnosis.

[0073] (2) Differences in optical radiation characteristics: Under different process conditions, the optical radiation characteristics generated by laser welding vary significantly. For example, fluctuations in laser power may cause changes in the intensity of the optical radiation signal, and different overlapping materials may affect the characteristics of back reflection. These factors cause the characteristic interval of the photoelectric signal defined as a good weld to change under different process conditions, thereby affecting the stability and accuracy of quality diagnosis.

[0074] (3) Lack of adaptability: Existing diagnostic methods usually require re-collection of a large amount of data to train the model under each process condition, which makes it difficult for them to meet the needs of industrial applications such as busbar laser welding, which require high efficiency and high reliability.

[0075] In order to fill this technical gap, the present invention proposes an innovative online diagnosis method and system for bus laser welding cold weld defects based on photoelectric signals. The method collects photoelectric signals in the laser welding process in real time, and uses a feature definition method based on reference values ​​to process and extract features from the signal data. Specifically, the feature definition method based on reference values ​​significantly improves the accuracy and robustness of the welding quality diagnosis model through multi-band data segmentation and signal difference coefficient calculation. In addition, the present invention also constructs a cold weld defect diagnosis model based on the random forest algorithm, which realizes efficient and accurate diagnosis of bus laser welding cold weld defects under different process conditions.

[0076] The main innovations and advantages are reflected in the following aspects:

[0077] Feature definition method based on reference value: The feature definition method based on reference value proposed in this invention overcomes the problem of signal feature interval variation caused by process parameter differences in traditional methods by segmenting multi-band data and calculating signal difference coefficients. Compared with the traditional feature definition method based on the signal itself, the feature definition method based on reference value significantly improves the accuracy and robustness of quality diagnosis, and can achieve stable and efficient defect diagnosis under various process conditions.

[0078] Defect diagnosis model based on random forest: The random forest model constructed by the present invention has an average accuracy rate of 99.20% in detecting cold weld defects in 1427 weld samples under different process conditions. This high accuracy rate ensures a reliable assessment of welding quality and provides a feasible diagnostic tool for industrial sites.

[0079] Defect inducing factor diagnosis: This method can not only accurately detect cold solder joint defects, but also further diagnose the inducing factors of cold solder joint defects, such as laser power fluctuation, focus shift, surface contamination and unstable joint gap. Experimental results show that the average accuracy of the model in diagnosing these inducing factors is 98.75%, providing more in-depth fault diagnosis support for production sites.

[0080] Through the above innovations, the technical solution of the present invention not only fills the current technical gap in the field of laser welding quality diagnosis at home and abroad, but also provides an intelligent, accurate and efficient solution for busbar and other high-demand welding quality control. Especially in the application of electric vehicles, power systems and new energy industries, this technology can greatly improve the quality control level of the welding process, reduce production costs, improve product safety and reliability, and provide strong support for technological progress and industrial upgrading in related industries.

[0081] Fourth, in the field of laser welding, especially in high-precision welding applications, real-time detection of welding quality and accurate diagnosis of cold weld defects have always been major challenges faced by technicians and engineers. Despite the continuous development of modern welding technology, there is still no mature and universal solution for how to achieve intelligent, accurate and real-time diagnosis of welding quality under complex process conditions.

[0082] (1) Difficulty in diagnosing welding quality

[0083] During the welding process, various factors such as laser power fluctuation, surface contamination, fixture instability, laser focus shift, etc. will affect the quality of the final weld. Although traditional welding quality inspection methods, such as image-based visual inspection and infrared thermal imaging inspection, have been applied to the quality control of the welding process to a certain extent, most of these methods have the following problems:

[0084] Reliance on manual experience: Traditional detection methods often rely on the experience of operators, with low levels of automation and intelligence, making it difficult to achieve efficient and stable quality control in large-scale production.

[0085] Affected by process conditions: During laser welding, changes in process parameters (such as laser power, focus position, material type, etc.) will cause significant changes in the characteristics of the photoelectric signal generated during welding. Traditional methods often require retraining the model under different process conditions, which greatly increases time cost and technical difficulty.

[0086] Lack of real-time performance: Most existing quality inspection methods are offline inspections, which cannot achieve real-time, online defect detection and quality assessment in the actual production process.

[0087] These problems have long plagued the industry, especially in high-precision and high-efficiency welding applications, where the accuracy and timeliness of welding quality detection are crucial to product quality control, production efficiency and equipment safety. Although researchers have made many attempts in this field, they have not been able to overcome the above difficulties and provide a solution that can adapt to complex process conditions and achieve efficient and intelligent diagnosis.

[0088] The technical solution of the present invention solves the above-mentioned problems through an intelligent diagnosis method based on photoelectric signals and machine learning. Specifically, the online diagnosis method and system for bus laser welding cold weld defects based on photoelectric signals proposed by the present invention introduces a feature definition method based on reference values, and effectively improves the accuracy and robustness of welding defect diagnosis through multi-band data segmentation and signal difference coefficient calculation, breaking through the limitations of traditional threshold judgment methods and successfully overcoming the problem of decreased diagnostic accuracy caused by process changes.

[0089] The core innovation of the present invention is:

[0090] Intelligent and efficient diagnosis: Through machine learning algorithms and combined with the characteristics of photoelectric signals, a welding quality diagnosis model based on random forests was established to achieve real-time and intelligent detection of welding defects. Compared with the traditional manual threshold setting method, this method has higher accuracy, stability and automation level in welding quality identification and defect detection.

[0091] Strong cross-process adaptability: By introducing a feature definition method based on reference values, the present invention can adapt to welding processes under different process conditions, overcoming the problem of reduced model accuracy in traditional methods when faced with complex factors such as different laser powers, surface contamination, and joint gaps. This method can perform automated diagnosis under a variety of welding conditions without the need to re-collect a large amount of data for training, thereby greatly improving diagnostic efficiency.

[0092] Real-time online detection: The present invention collects photoelectric signals in the laser welding process in real time and uses efficient algorithms to quickly analyze and diagnose, thus solving the problem of the inability to perform real-time, online welding quality detection in the prior art. This innovation enables the present invention to be directly applied to the production line, realize dynamic monitoring of welding quality, quickly locate defects and adjust the production process in a timely manner, greatly improving production efficiency.

[0093] High accuracy and robustness: The cold weld defect diagnosis model proposed in this invention has an average accuracy rate of 99.20% in cold weld defect detection of 1427 weld samples under different process conditions, and the diagnostic accuracy rate for defect-inducing factors (such as laser power fluctuation, focus shift, surface contamination, etc.) is 98.75%. This high accuracy rate enables the present invention to effectively solve the problem of insufficient diagnostic accuracy of traditional methods and meet the needs of high-precision welding quality control.

[0094] Therefore, the technical solution of the present invention not only successfully breaks through the limitations of the traditional photoelectric signal welding quality diagnosis method and fills the technical gap in this field, but also provides an efficient, intelligent and adaptable solution for welding quality control. It solves the technical problem that people have long been eager to solve, that is, how to use photoelectric signals to realize real-time and accurate online diagnosis of cold soldering defects in the welding process under complex process conditions, and on this basis, further diagnose the inducing factors of the defects. BRIEF DESCRIPTION OF THE DRAWINGS

[0095] Figure 1 It is a flow chart of an online diagnosis method for busbar laser welding cold soldering defects based on photoelectric signals provided by an embodiment of the present invention;

[0096] Figure 2 It is a structural diagram of an online diagnosis system for busbar laser welding cold welding defects based on photoelectric signals provided by an embodiment of the present invention;

[0097] Figure 3 is a schematic diagram of a laser welding and photoelectric monitoring system provided by an embodiment of the present invention;

[0098] Figure 4 is a schematic diagram of a signal source provided by an embodiment of the present invention;

[0099] Figure 5 is a schematic diagram of data segmentation provided by an embodiment of the present invention;

[0100] Figure 6 is a schematic diagram of feature extraction provided by an embodiment of the present invention;

[0101] Figure 7 It is a schematic diagram of random forest model connection provided by an embodiment of the present invention;

[0102] Figure 8 is a schematic diagram of obtaining training data based on defect induction provided by an embodiment of the present invention;

[0103] Fig. 9 It is a schematic diagram of a signal difference coefficient characteristic matrix provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0104] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0105] like Figure 1 As shown, the online diagnosis method of bus laser welding cold welding defects based on photoelectric signals includes:

[0106] S1, raw photoelectric signal data processing;

[0107] S2, feature definition based on reference values;

[0108] S3, build a quality diagnosis model based on random forest.

[0109] The photoelectric signal generated during the bus laser welding process is collected in real time by a high-speed sensor, and the signal is subjected to noise reduction, smoothing and normalization to eliminate environmental interference and obtain stable and consistent photoelectric signal data;

[0110] Based on the processed photoelectric signal, key characteristic parameters are extracted, including light intensity fluctuation, signal mean and frequency component. The extracted features are compared with standard reference values ​​to define specific characteristic indicators, such as signal amplitude deviation and penetration signal non-uniformity.

[0111] The quality diagnosis model is trained using the random forest algorithm. By inputting the characteristic data of the photoelectric signal, the welding status is classified as "normal" or "weak welding", and the mapping relationship between the characteristics and the welding quality is learned through the model.

[0112] Input the photoelectric signals collected in real time into the constructed quality diagnosis model to detect the welding quality in real time;

[0113] When a cold weld defect is detected, an adjustment signal is immediately sent to the welding equipment to optimize the laser welding parameters, including laser power, scanning speed or other key parameters, to ensure the stability and reliability of the welding quality.

[0114] like Figure 2 As shown, the bus laser welding cold soldering defect online diagnosis system based on photoelectric signals includes:

[0115] A data processing module, used for processing raw photoelectric signal data;

[0116] A feature definition module, used for defining features based on reference values;

[0117] The quality diagnosis model building module is used to build a quality diagnosis model based on random forest.

[0118] During the busbar laser welding process, the welding spot will generate continuous photoelectric signals, which contain key information about the welding quality. The data processing module collects the original photoelectric signals in real time through high-speed sensors and pre-processes the signals. Pre-processing includes signal noise reduction, smoothing and normalization to eliminate the influence of environmental interference and noise and ensure the stability and consistency of the signal. This process provides high-quality data input for subsequent feature extraction and diagnostic analysis.

[0119] The feature definition module extracts key parameters related to welding quality, such as light intensity fluctuation, signal mean, frequency components, etc., by analyzing the waveform characteristics of the photoelectric signal. These features are further defined by comparing them with standard reference values ​​(such as signal amplitude deviation, penetration signal non-uniformity, etc.). These reference values ​​are established from historical data of normal welding signals to ensure that the feature definition can accurately reflect the occurrence of cold weld defects.

[0120] Random forest is a machine learning algorithm based on multiple decision trees, which is used to handle the classification problem of multidimensional data. The quality diagnosis model building module uses a large amount of photoelectric signal feature data for training, and divides the welding signal into two states: "normal" or "weak solder joint". Through the input of feature data, the random forest model automatically learns the mapping relationship between features and welding quality, and improves the ability to identify weak solder joint defects. After training, the model can receive photoelectric signals in real time and output diagnostic results.

[0121] The entire system operates in online mode, collecting welding photoelectric signals in real time and transmitting the signals to the diagnostic module for analysis. If the diagnostic model detects a cold weld defect, the system will immediately send out an alarm signal and feed the information back to the welding equipment, prompting the operator or automatically adjusting the welding parameters (such as laser power, scanning speed, etc.). This closed-loop control mechanism ensures real-time monitoring and rapid adjustment of welding quality, reduces the incidence of cold weld defects, and improves the efficiency of the production line and product reliability.

[0122] 1. The device provided by the present invention

[0123] Fiber laser, 2D galvanometer welding head, optical fiber, photoelectric sensor, signal acquisition and analysis device and host computer.

[0124] Among them, the laser welding system combines the mechanical mechanism with the composite motion mode of the galvanometer motor, which can achieve high-precision processing of large areas and arbitrary positions. The galvanometer control software provides highly flexible laser parameter editing functions, including settings such as swing trajectory and power, which can optimize the heat input of the weld in time and space scales to meet different welding requirements.

[0125] In order to obtain photoelectric signals that can effectively characterize the welding quality, photodiodes are used to monitor the optical radiation signals of three bands, namely near-infrared light, infrared light and blue-violet light, during the laser welding process. These signals come from laser reflection, molten pool thermal radiation and plasma plume during the laser welding process. Optical signals of different wavelengths will be generated during the welding process. The deflection lens in the galvanometer system is used to transmit the collected optical signals to the fully reflective and semi-transparent 45-degree lens. The 45-degree lens splits the irradiated optical signals and uses filters to filter out the optical signals of wavelengths that do not need to be detected. Then, optical components such as focusing lenses are used to adjust the input, and photoelectric signals of specific wavelengths are collected by photodiodes to obtain voltage information reflecting the intensity of the optical signals.

[0126] 2 Methods

[0127] 2.1 Raw photoelectric signal data processing

[0128] During each laser welding process, the photodiode records a sequence of sampling points of light signals in three bands. Each sampling point contains information about time, band, and its corresponding signal strength. In order to extract valid data, the invalid part where no welding occurs is first removed by setting a suitable threshold, and the time point when welding starts is calculated to determine the valid interval of the signal.

[0129] In order to eliminate the high-frequency noise interference in the original signal, a low-pass filter is used to process the signal. Low-pass filtering can remove the high-frequency components in the signal while retaining the original characteristics of the low frequency, thereby smoothing the signal curve. The function of the low-pass filter is to filter the signal through formula 1.

[0130]

[0131] Among them, x(t) is the original signal, h(t) is the impulse response function of the filter, and common low-pass filters include ideal low-pass filter, Butterworth low-pass filter, etc.

[0132]

[0133] The frequency response of an ideal low-pass filter is shown in Equation 2. Where f_c is the cutoff frequency of the filter, indicating the highest frequency passed. After filtering, a three-band "time-signal intensity" graph is plotted for further analysis.

[0134] 2.2 Feature definition method based on reference value

[0135] The present invention proposes a feature definition method based on reference values. This method calculates features by taking the quality assessment template under specific process conditions as a reference, thereby improving the reliability of the quality diagnosis model and enabling it to adapt to the welding process diagnosis under different process conditions. In order to fully reflect the laser welding process and avoid missing local defects, the present invention adopts a multi-band data segmentation method and calculates the signal difference coefficient of each band. Among them, the signal difference coefficient represents the difference ratio between the signal and the reference in each data segment, reflecting information such as the degree and direction of the difference between the signal and the reference. The signal difference coefficients are arranged into a signal difference coefficient-feature matrix according to the specified positions as the features of the input model, which not only increases the interpretability of the model, but also provides support for further subdivision of defects. The specific calculation process of feature extraction is as follows.

[0136] The reference curve is XY, the measured sample curve is xy, and the total number of sample points is n.

[0137] First, the quality assessment template is calculated. Each sample corresponds to a good weld test sample group containing at least 15 samples under its process conditions. The average value of the intensity of the photoelectric signal data of all samples in the good weld test sample group within the same band range and at the same time point is calculated to generate a reference for the three bands of the sample.

[0138] Secondly, perform multi-band data segmentation. First, the photoelectric signals of the reference curve and the measured sample curve need to be aligned to ensure that they occur in the same time series, and then the sample data is segmented. Specifically, the photoelectric signal data is divided into 50 segments by time, and the basic sampling points a and remainder b of each segment are calculated to satisfy Formula 3.

[0139] n=50a+b, 0≤b<50 (3)

[0140] According to this segmentation rule, each of the first 49 segments contains a sampling points, and the number of sampling points in the 50th segment is a + b. If n is divisible by 50 (ie, b = 0), then all 50 segments contain a sampling points.

[0141] Next, calculate the signal difference coefficient in each data segment. Take the signal difference coefficient of the i-th data segment in the first band as an example, its position in the matrix is ​​(1, i). The calculation rule of the signal difference coefficient is as follows: calculate the ratio of the vertical coordinate difference of all sampling points between the measured sample curve and the reference curve in the data segment to the vertical coordinate of the corresponding sampling point of the reference curve, and take the average of these ratios. Assuming that the time coordinate value of the sampling point contained in this segment of data is from T1 to T2, the calculation of the signal difference coefficient satisfies Formula 4.

[0142]

[0143] Finally, a complete signal difference coefficient feature matrix is ​​obtained. The signal difference coefficients of all data segments in the three bands of the experimental sample are calculated to generate a 3x50 signal difference coefficient feature matrix. The signal difference coefficient feature matrix describes the degree of difference between the measured sample curve and the reference curve, covering information such as the amplitude, direction, time and band of the difference, and unifies the judgment criteria for good welds under different welding conditions. This method ensures that the entire welding process, including local cold weld defects, is captured in an accurate and adaptable manner to different welding conditions, providing important basic data for subsequent machine learning model training and welding quality diagnosis.

[0144] 2.3 Building a quality diagnosis model based on random forest

[0145] Dataset construction: Multiple sub-datasets are generated from the original training set through Bootstrap sampling. Each sub-dataset consists of some samples in the original dataset, and some samples may be repeated due to sampling with replacement. Assume that the training set D = {x_1, x_2, ..., x_N}, where xi represents the training sample and N is the number of samples. In the Bootstrap sampling process, N samples are randomly selected from the training set D each time, and repeated samples may appear.

[0146] Decision tree training: For each sub-dataset D1 generated by Bootstrap sampling, a decision tree is constructed. At each node of the tree, a subset of features is randomly selected (random selection of features) to find the optimal splitting feature. The node is recursively split until the stopping criterion (such as the depth of the tree or the number of samples in the leaf node) is met. The construction of the decision tree relies on Information Gain (IG) and Gini Index to measure the quality of the split.

[0147] Information gain is a measure of the improvement in class purity after splitting a data set by a certain feature. Assuming that feature A is selected for splitting at node n, the information gain calculation formula is:

[0148]

[0149] Among them, H(D) is the entropy of the data set D, which represents the uncertainty of the data set:

[0150]

[0151] Among them, p i is the probability of category i in the data set, m is the total number of categories; D v It is the data subset when feature A takes the value v.

[0152] The Gini index is used to measure the impurity of a data set. The smaller the Gini index, the higher the purity of the data set. The calculation formula of the Gini index is:

[0153]

[0154] Where pi is the probability of belonging to category i in dataset D.

[0155] Integration and voting: After each decision tree is trained, the prediction results of all decision trees are integrated by majority voting. In the test phase, for the input sample x, each decision tree Ti will output a category prediction Ci. The final prediction category C is determined by voting:

[0156] C=mode(C1,C2,...,C N ) (8)

[0157] Among them, N is the number of decision trees in the random forest, and mode means selecting the category with the highest frequency.

[0158] Output: As the classification results of all decision trees are integrated, the final classification result is the output of the random forest. This integration method can significantly improve the classification performance and reduce the bias and variance that a single decision tree may produce on the training set.

[0159] Among them, N is the number of decision trees in the random forest, and mode means selecting the category with the highest frequency.

[0160] In order to optimize the performance of the random forest model, the present invention tunes two hyperparameters, the number of decision trees and the minimum number of leaves. The tuning process adopts a strategy of grid search combined with Bayesian optimization.

[0161] First, we systematically explore the hyperparameter space through grid search to provide initial results for subsequent Bayesian optimization. The hyperparameter ranges of the grid search are set as follows: minimum number of leaves (1, 3, 5, 7, 9) and number of decision trees (50, 100, 150, 200, 250). Performance evaluation is performed through 5-fold cross validation on the data set. The optimal hyperparameter combination obtained by grid search is 1 minimum number of leaves and 100 decision trees.

[0162] On this basis, in order to further optimize the model performance, Bayesian optimization is used to tune within a finer range of hyperparameters. Specifically, the minimum number of leaves ranges from 1 to 6, and the number of decision trees ranges from 50 to 150. Bayesian optimization iteratively adjusts parameters based on the results of the previous round and eventually converges to the global optimal hyperparameter combination. The best observed parameter combination is a minimum number of leaves of 1 and a number of decision trees of 107. The best parameter combination obtained based on the surrogate model is a minimum number of leaves of 1 and a number of decision trees of 143. The final hyperparameter combination significantly improves the accuracy and robustness of the model, providing reliable support for diagnostic tasks.

[0163] The factors causing cold solder joint defects are classified into: laser power fluctuation, unstable defocus, surface contamination, and changes in the overlap gap. By adjusting the welding parameters and conditions, several defect types are induced to simulate the situations that may cause cold solder joints in actual production.

[0164] (1) Insufficient laser power: By reducing the laser power to a level lower than that required for normal welding, the situation of insufficient laser power is simulated. In the actual production process, insufficient laser power may be caused by equipment aging, power control system failure, etc., resulting in cold weld defects. In order to verify the detection capability of this scheme for local cold welds, in some samples, insufficient laser power is set only in the local welding process, and the occurrence time is less than 0.25 seconds.

[0165] (2) Defocus change: By adjusting the distance between the welding head and the workpiece, the defocus change condition is set. In the actual production process, the distance between the workpiece and the welding head may change due to insufficient repeated positioning accuracy, equipment calibration problems, etc., resulting in unstable defocus and causing cold welding defects.

[0166] (3) Surface contamination: Glue is dripped on the welding track to simulate the presence of contaminants on the workpiece surface. In the actual busbar production process, contaminants such as grease, dust or glue residue may appear on the workpiece surface due to processing residues, poor storage environment, etc., which will affect the welding quality and cause cold welding defects.

[0167] (4) Overlap gap change: By increasing the welding overlap gap, the situation of excessive overlap gap is simulated. In actual production, the overlap gap may be too large due to reasons such as loose fixtures and errors in mechanical parts, which will lead to cold welding defects.

[0168] Obtaining defect-induced sample data to train the model can enable the model to diagnose the factors causing cold soldering defects.

[0169] This method can not only detect whether there are cold weld defects in bus laser welding, but also further diagnose the inducing factors of cold weld defects, which provides effective support for quickly locating the source of the problem and taking targeted measures in industrial production.

[0170] 1 Experiment

[0171] A 3000W fiber laser and a 2D galvanometer welding head were used for laser welding experiments. The experimental materials included two groups of busbar samples with different thicknesses. The upper layer of the first group of materials was 1.5mm thick 1060 aluminum alloy, and the lower layer was 2.5mm thick 6061 aluminum alloy; the upper layer of the second group of materials was 1mm thick 1060 aluminum alloy, and the lower layer was 3mm thick 6061 aluminum alloy. Photodiodes were used to monitor the optical radiation signals of three bands of near-infrared light, infrared light, and blue-violet light during the laser welding process.

[0172] The laser lap welding process is used, combined with the galvanometer welding head to realize the linear swing and linear motion superposition of the laser beam. The upper and lower workpieces are fixed by a fixture, and gaskets of different thicknesses are added between the workpieces to adjust the lap gap. In order to study the detection method of cold weld defects in bus laser welding, 7 groups of experiments were set up, each group containing good weld samples and weld samples simulating cold weld defects. The good weld process conditions (welding trajectory, welding parameters, experimental materials, etc.) in each group are different. The samples are generated by defect induction method. Specific types of defects (laser power fluctuations, unstable defocus, surface contamination, and lap gap changes) are artificially introduced. In different experimental groups, the defect samples include the above-mentioned trap induction methods, and a total of 1427 weld samples were collected. Each sample records information such as photoelectric signal data, process parameters, and weld status to form a complete sample set. The signal data collected in real time using a photodiode will be used for subsequent feature extraction and machine learning model training.

[0173] 2 Effects

[0174] In order to simulate the number of samples in the normal weld reference experimental sample group in the actual production process, in addition to using the complete experimental sample set as a reference, we also re-selected some normal weld reference experimental samples. This not only helps to verify the impact of reducing the number of reference samples on the accuracy of defect detection, but also expands the number of data samples. In previous experiments, 1427 weld samples and their corresponding photoelectric signal raw data were obtained. After data processing, some samples used different reference curves, and finally formed 2373 groups of data samples for training machine learning models. After the data processing is completed, we mark the status of the sample data as positive samples (representing normal welds) and negative samples (representing the presence of virtual weld defects) according to whether there are welding defects. These 2373 groups of data contain 1283 positive samples and 1090 negative samples.

[0175] In order to more comprehensively evaluate the generalization ability and stability of the model, the data set was randomly shuffled 20 times, 20 rounds of training were performed, and 5-fold cross validation was performed in each round, and the average performance indicators were finally taken. In the task of cold soldering defect detection, the average accuracy, precision, recall rate and F1 value were 99.20%, 98.60%, 98.28% and 98.93% respectively.

[0176] In the diagnosis task of different inducing factors of cold solder joint defects, 318 groups of cold solder joint samples were selected and marked as four inducing factors: laser power fluctuation (72 groups), defocus instability (84 groups), surface contamination (89 groups) and overlap gap change (74 groups), and each group contained samples of process conditions. The average accuracy, precision, recall and F1 value were 98.75%, 98.43%, 98.95% and 98.68% respectively.

[0177] An application embodiment of the present invention provides a computer device, which includes a memory and a processor. The memory stores a computer program. When the computer program is executed by the processor, the processor executes the steps of the method.

[0178] An application embodiment of the present invention provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the processor executes the steps of the method.

[0179] An application embodiment of the present invention provides an information data processing terminal, which includes a system.

[0180] The present invention is based on the machine learning online diagnosis method of cold solder joint defects driven by photoelectric signals, which has a wide range of applications and huge market potential. Its core advantage is that it can provide real-time and accurate welding quality diagnosis and can adapt to welding processes under different process conditions. Products and solutions based on this technology can be widely used in the following fields:

[0181] 1. Electric vehicle industry

[0182] With the promotion of global energy transformation and environmental protection policies, the market demand for electric vehicles (EVs) has grown rapidly, especially in terms of reducing greenhouse gas emissions and improving energy efficiency. EVs have become an important part of the future transportation sector. In the production process of electric vehicles, the welding quality of battery packs is crucial to their performance and safety. Each electric vehicle battery pack usually requires tens of thousands of batteries to be welded, which places extremely high demands on the accurate control of welding quality.

[0183] The cold soldering defect diagnosis technology of the present invention can provide an efficient and accurate quality control solution for the production of electric vehicle battery packs. By real-time monitoring of the quality of the welding process, cold soldering defects and their inducing factors (such as laser power fluctuations, focus shifts, etc.) are detected in a timely manner to ensure that the welding quality reaches the highest standards. This technology can not only effectively improve the safety of the battery pack, but also reduce rework and waste caused by poor welding, thereby reducing production costs.

[0184] 2. New energy industry (including solar energy, wind energy, etc.)

[0185] In the new energy industry, especially in the production of solar photovoltaic and wind power equipment, the quality of electrical connections directly affects the efficiency and long-term stability of the system. In the manufacturing process of solar panels and wind turbines, the welding quality of cells, conductive strips and wiring parts is crucial. For the welding quality of these products, traditional detection methods are often difficult to adapt to the high-speed and precise production requirements.

[0186] The technology of the present invention can realize real-time online diagnosis of welding quality in the production process of new energy products, accurately identify cold welding defects, thereby improving the working efficiency and reliability of the equipment. Through efficient defect detection and diagnosis, the defect rate in the production process can be greatly reduced, ensuring that the new energy equipment is stable and reliable during operation.

[0187] 3. Power system equipment manufacturing

[0188] In the power system, busbars are a key conductive component widely used in power distribution systems, substations and other facilities. The welding quality of busbars is directly related to the safety and stability of power equipment. Due to the complex welding environment in the power system, traditional quality inspection methods are difficult to diagnose welding defects in real time and accurately.

[0189] The technical solution of the present invention can be applied to the welding quality diagnosis of power equipment, especially in the manufacture of busbars and other key components, to monitor the welding quality in real time through photoelectric signals to ensure that defects in the welding process are identified and corrected in a timely manner. This not only improves the operational reliability of power equipment, but also reduces the risk of equipment failure caused by welding defects.

[0190] 4. Industrial automation and robotic welding

[0191] Industrial automation and robotic welding systems are widely used in many industries such as automobiles, home appliances, and metal processing. Welding is an important part of the automated production line, but how to achieve real-time quality monitoring and defect diagnosis in the high-speed and high-precision welding process has always been a technical challenge.

[0192] The present invention provides an efficient quality monitoring and defect diagnosis tool for industrial automation production lines through real-time detection based on photoelectric signals and diagnostic capabilities of machine learning algorithms. Applied in automated welding systems, it can achieve precise control of each welding process, improve production efficiency and quality stability.

[0193] The present invention realizes real-time online diagnosis of cold welding defects during laser welding by means of a photoelectric signal monitoring method. The experiment uses a 3000W fiber laser and a 2D galvanometer welding head, equipped with multiple photodiodes, to monitor photoelectric signals of different bands during laser welding in real time. The welding materials used in the experiment are aluminum alloys of different thicknesses, namely 1060 aluminum alloy and 6061 aluminum alloy.

[0194] In the experiment, the photodiode monitored the light radiation signals in three bands: near-infrared light, infrared light and blue-violet light. The signal data was collected and used to train the quality diagnosis model. The signal features were extracted by the feature definition method based on the reference value, and a quality diagnosis model based on random forest was constructed.

[0195] The present invention adopts a laser lap welding process and combines a galvanometer welding head to realize the linear swing and linear motion superposition of the laser beam. The upper and lower workpieces are fixed by a clamp, and gaskets of different thicknesses are added between the workpieces to adjust the lap gap. In order to study the detection method of cold welding defects in bus laser welding, 7 groups of experiments were set up, each group containing good weld samples and weld samples simulating cold welding defects. The good weld process conditions (welding trajectory, welding parameters, experimental materials, etc.) in each group are different.

[0196] Since defects in industrial applications occur infrequently and only under certain conditions, it is difficult to collect real defect samples. To ensure that the machine learning model can effectively learn defect characteristics, the present invention uses a defect induction method to generate samples. By artificially introducing specific types of defects, the diversity and comprehensiveness of defect samples in the training data can be ensured, the generalization ability of the model can be improved, and the type, size and location of the defects can be accurately controlled, thereby improving the accuracy and reliability of the model.

[0197] This paper uses a stratified 5-fold cross-validation method to validate the model. In order to prove the advantages of the random forest model in this task, we selected several other models as the control group, used the same data set and evaluation method for validation, and then compared the performance of different models. This paper compares the back propagation neural network (BP) model, the long short-term memory network (LSTM) model, and the KNN model.

[0198] The experimental results show that the average accuracy of the cold weld defect detection using the method of the present invention in 1427 weld samples is 99.20%. In addition, the model can effectively diagnose the inducing factors of cold weld defects in the welding process, with a diagnostic accuracy of 98.75%. These experimental results prove the efficient and accurate detection ability of the present invention for welding defects under different process conditions.

[0199] Table 1 Experimental parameters (a set of good weld parameters)

[0200]

[0201] Table 2 Model training results for cold soldering defect detection

[0202]

[0203]

[0204] Table 3 Model training results for diagnosis of cold solder joint inducing factors

[0205]

[0206] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware part can be implemented using dedicated logic; the software part can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. It can be understood by a person of ordinary skill in the art that the above-mentioned devices and methods can be implemented using computer executable instructions and / or contained in a processor control code, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. Such code is provided on the carrier medium. The device and its modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, and can also be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0207] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principle of the present invention should be covered by the protection scope of the present invention.

Claims

1. An online diagnosis method for busbar laser welding cold welding defects based on photoelectric signals, characterized in that: include: S1, collects the photoelectric signals generated during the bus laser welding process in real time through high-speed sensors, and performs noise reduction, smoothing and normalization on the signals to eliminate environmental interference and obtain stable and consistent signal data; S2, based on the processed photoelectric signal, extract key characteristic parameters, compare the characteristics with standard reference values, define specific characteristic indicators, and use the random forest algorithm to build a quality diagnosis model to classify the welding status as "normal" or "cold welding"; S3, inputs the real-time collected photoelectric signal into the quality diagnosis model for welding quality detection. When a cold soldering defect is detected, an adjustment signal is immediately sent to the welding equipment to optimize the welding parameters and realize real-time quality optimization and control.

2. The on-line diagnosis method for busbar laser welding cold solder joint defects based on photoelectric signals according to claim 1, characterized in that: Step S1 includes: during each laser welding process, using a photodiode to record a sequence of optical signal sampling points in multiple bands, each sampling point containing time, band and signal strength information; removing invalid parts by setting a threshold, calculating the start time of welding to determine the valid interval of the signal, and using a low-pass filtering method to smooth the signal curve and eliminate high-frequency noise interference.

3. The on-line diagnosis method for busbar laser welding cold solder joint defects based on photoelectric signals as claimed in claim 1, characterized in that: Step S2 includes: calculating the optical signal characteristics by referring to the quality assessment template under specific process conditions, and using a multi-band data segmentation method to calculate the signal difference coefficient of each band; arranging the signal difference coefficients into a feature matrix and using them as features of the input model, thereby improving the adaptability of the diagnostic model under different process conditions.

4. The on-line diagnosis method for busbar laser welding cold solder joint defects based on photoelectric signals as claimed in claim 1, characterized in that: Step S2 further includes: calculating the signal difference coefficient between the reference curve and the measured sample curve; calculating the signal difference ratio of each segment of the photoelectric signal data by time segment, and generating a multi-band feature matrix; the matrix contains the amplitude, direction and time information of the difference, providing comprehensive basic data for model training.

5. The on-line diagnosis method for busbar laser welding cold solder joint defects based on photoelectric signals as claimed in claim 1, characterized in that: Step S3 includes: using Bootstrap sampling technology to generate multiple sub-datasets from the original data set, and training a decision tree based on each sub-dataset; selecting the optimal split feature by measuring information gain and Gini index, recursively constructing a decision tree, and forming a random forest model.

6. The on-line diagnosis method for busbar laser welding cold solder joint defects based on photoelectric signals according to claim 1, characterized in that: Step S3 includes: integrating the classification results of all decision trees by majority voting; in the testing phase, outputting the final classification result according to the voting results of all decision trees to improve classification accuracy and reduce model bias.

7. The on-line diagnosis method for busbar laser welding cold solder joint defects based on photoelectric signals according to claim 1, characterized in that: The number of decision trees and the minimum number of leaves of the random forest model were tuned, and the optimal hyperparameter combination was determined by grid search combined with Bayesian optimization to improve the accuracy and robustness of the model.

8. An online diagnosis system for busbar laser welding cold soldering defects based on photoelectric signals, which implements the online diagnosis method for busbar laser welding cold soldering defects based on photoelectric signals as claimed in any one of claims 1 to 7, characterized in that: include: A data processing module, used for processing raw photoelectric signal data; A feature definition module, used for defining features based on reference values; The quality diagnosis model building module is used to build a quality diagnosis model based on random forest.

9. A computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to execute the steps of the online diagnosis method for busbar laser welding cold weld defects based on photoelectric signals as described in any one of claims 1 to 6.

10. An information data processing terminal, comprising the on-line diagnosis system for busbar laser welding cold solder joint defects based on photoelectric signals as claimed in claim 7.

Images