A spot welding machine welding head self-adjusting method and system based on CCD-AI detection

Abstract

The application discloses a kind of based on CCD-AI detection's spot welding machine welding head self-regulation method and system, comprising the following steps: S00: image acquisition and processing;S10: state identification and classification;S20: adjustment decision generation;S30: instruction execution and adjustment;S40: trend prediction and forward-looking adjustment;Step S50: self-learning and rule optimization;To acquire spot image by CCD camera, in combination with deep learning model (such as ResNet-50) is intelligently classified to spot state, can identify including not welding, false welding, weld mark size anomaly and a variety of defects, and according to classification setting and spot welding machine cooperation spot state instruction parameter mapping rule base, according to the welding point state identified automatically generates welding parameter adjustment instruction (such as welding voltage), let CCD-AI vision detection module and spot welding machine form self-regulation mode between, to improve welding quality, reduce and eliminate the problem of relying on artificial debugging parameter error and adjustment not in time.

Description

Technical Field

[0001] This invention relates to the field of weld joint inspection, and in particular to a self-adjusting method and system for a spot welding machine head based on CCD-AI inspection. Background Technology

[0002] In the soldering process of electronic products (such as soldering the voice coil and FPC leads in a speaker), the solder joint quality deteriorates due to wear or contamination of the welding head caused by prolonged use of the spot welding machine. This can result in issues like incomplete solder joints, excessively large or small solder marks. To address this problem, conventional methods for monitoring solder joint condition involve using a vision camera to collect solder joint information and employing image processing algorithms to analyze the product's surface features. Detection is achieved by comparing the product image with a standard model. For example, the ratio can be calculated by measuring the actual distance to the calibration object and the corresponding pixel value in the image, allowing for timely detection of defective products. Defective products can then be addressed by manually adjusting the spot welding machine voltage: increasing the voltage when the solder mark is small, and adjusting it when the solder mark is large. Reducing the voltage of the spot welding machine creates an open-loop process of detection, manual judgment, and manual adjustment. In this case, existing technologies, based on deep learning, have developed online detection and closed-loop control methods encompassing image acquisition, feature extraction / defect recognition, decision generation, and parameter adjustment. For example, patent CN120269548A discloses a deep learning-based intelligent welding robot control method and system; patent CN120084806A discloses a visual inspection method, device, and computer equipment for reflow soldering quality; and patent CN120002132A discloses… A visual sensing-based arc welding robot welding monitoring system and method are presented. All three utilize a closed-loop feedback control process during welding, achieving the goal of reducing manual intervention and improving welding quality and efficiency. However, in the case of smaller electronic products (such as the welding of voice coils in speakers and lead wires in FPCs as mentioned above), the spot welding (resistance welding) process still suffers from weak identification capabilities for complex defects. In particular, when the welding head is worn or complex working conditions change (such as scratches, stains, etc.), the defect detection accuracy is low and the miss rate is high, making it difficult to adjust the corresponding parameters of the spot welding machine in a timely and effective manner. Summary of the Invention

[0003] In order to overcome the shortcomings of the prior art, the technical problem to be solved by the present invention is to propose a self-adjusting method and system for spot welding machine welding head based on CCD-AI detection. The CCD-AI vision detection system monitors the welding marks of the spot welding machine, and then provides real-time feedback and automatically adjusts the response parameters of the spot welding machine to optimize the welding effect of the spot welding machine welding head.

[0004] To achieve this objective, the present invention adopts the following technical solution:

[0005] This invention provides a self-adjusting method for a spot welding machine welding head based on CCD-AI detection, comprising the following steps:

[0006] S00: Image acquisition and processing. The CCD~AI vision inspection module acquires images of the weld points after welding by the spot welding machine and processes the images to extract the weld point area.

[0007] S10: Status recognition and classification. The CCD~AI vision detection module inputs the collected solder joint area image into the pre-trained spot welding deep learning model. The spot welding deep learning model outputs the status category label of the solder joint. The status category label includes normal, unwelded, cold solder joint and large solder mark.

[0008] S20: Adjustment decision generation. The CCD~AI vision detection module queries the preset weld point status~instruction parameter mapping rule library according to the status category label, generates parameter adjustment instructions for the spot welding machine and sends them to the logic control module. The weld point status~instruction parameter mapping rule library is set with the welding parameter adjustment value of the spot welding machine corresponding to the status category label of the weld point.

[0009] In step S20, before generating parameter adjustment instructions for the spot welding machine, the following steps are also included:

[0010] S21: Continuous state judgment, statistical analysis of the state identification results of N solder joints produced in continuous production;

[0011] S22: Trigger condition determination: The parameter adjustment command for the spot welding machine will be generated and an early warning will be triggered only when M consecutive abnormal weld points of the same type occur, or when the frequency of abnormal occurrence exceeds the preset threshold. M and N are preset positive integers, and M≤N.

[0012] In step S20, the solder joint status ~ instruction parameter mapping rule base is a quantization mapping base, which includes at least the following instructions:

[0013] The "not welded" status label is mapped to an instruction indicating that the spot welding machine's welding head is damaged.

[0014] The cold solder joint status label is mapped to an increase in soldering voltage. The instructions also include a small solder stamp status label mapped to an increase in soldering voltage. The instructions, in which ;

[0015] Large solder mark status label mapped to reduced soldering voltage The instructions;

[0016] The normal status label is further divided into an OK status label, an OK solder mark small label, and an OK solder mark large label. The OK status label is mapped to the spot welding machine voltage normal command, and the OK solder mark small label is mapped to the increase of welding voltage. The instruction that the OK soldering mark is mapped to an increased soldering voltage The instructions.

[0017] In step S20, the weld point status ~ instruction parameter mapping rule base is also equipped with a digital twin model. After generating parameter adjustment instructions for the spot welding machine, the parameter adjustment instructions and the current working condition are input into the digital twin model of the spot welding process for simulation. If the simulation result predicts that the weld point quality meets the requirements and there is no equipment risk, then it is confirmed to continue executing step S30; otherwise, the parameter adjustment instruction is corrected or canceled and an early warning is issued.

[0018] S30: Instruction execution and adjustment. The logic control module sends the parameter adjustment instruction to the controller of the spot welding machine, driving the spot welding machine to automatically adjust the welding parameters to compensate for wear of the welding head or changes in working conditions.

[0019] To further perform multimodal processing based on image features, the spot welding deep learning model identifies the weld point status by fusing image information acquired by the CCD~AI vision detection module with non-image information from at least one auxiliary sensor. The non-image information includes acoustic signals, electrical signals, or thermal signals during welding. The spot welding deep learning model outputs a weld point status category label and a corresponding comprehensive confidence score. Only when the comprehensive confidence score is higher than a preset threshold is a parameter adjustment instruction generated based on the status category label in step S20. According to the identified weld point status category label and the image features of the weld point area, the optimal coordinated adjustment amount of each parameter is obtained through a multi-parameter optimized spot welding deep learning model. The parameter adjustment instruction includes an instruction to coordinately adjust at least two parameters among welding voltage, electrode pressure, and welding time.

[0020] It also includes trend prediction and forward-looking adjustment steps, recording and analyzing feature data of historical solder joint images to construct a time-series trend of solder joint quality changes; based on the trend of changes, judging the wear trend of the solder head or the stability of the process through a prediction model; when the prediction result indicates that the risk of quality degradation exceeds a preset threshold, before the solder joint status is identified as abnormal, generating and executing a forward-looking preventive parameter fine-tuning instruction.

[0021] It also includes self-learning and rule optimization steps, continuously collecting weld point images and their status data after the execution of adjustment instructions to form a feedback dataset; based on the feedback dataset, incrementally training the spot welding deep learning model to optimize recognition accuracy, and / or adaptively adjusting the mapping relationship in the weld point status ~ instruction parameter mapping rule base to optimize the adjustment effect.

[0022] A self-adjusting system for a spot welding machine head based on CCD-AI detection, used to implement the self-adjusting method for a spot welding machine head based on CCD-AI detection as described above, includes the following modules:

[0023] The CCD-AI vision inspection module includes an industrial computer and an inspection camera for acquiring weld point images. The industrial computer has a built-in AI processing module for running a spot welding deep learning model and a weld point state-instruction parameter mapping rule library to perform state recognition and generate parameter adjustment instructions. A logic control module receives the parameter adjustment instructions and converts them into control signals. A spot welding machine module is controlled by the control signals to adjust its welding parameters. The logic control module is a PLC control system, which communicates with the CCD-AI vision inspection module and the spot welding machine module through an industrial communication protocol to parse the parameter adjustment instructions and generate specific control pulse sequences.

[0024] The beneficial effects of this invention are as follows:

[0025] (1) This case uses a CCD camera to collect weld point images and combines deep learning models (such as ResNet-50) to intelligently classify the weld point status. It can identify various defects, including no welding, cold welding, and abnormal weld stamp size. Based on the classification, a weld point status ~ instruction parameter mapping rule library is set to cooperate with the spot welding machine. Based on the identified weld point status, welding parameter adjustment instructions (such as welding voltage) are automatically generated, so that the CCD ~ AI vision detection module and the spot welding machine form a self-adjustment mode to improve welding quality and reduce and eliminate the problems of relying on manual adjustment of parameters and untimely adjustment.

[0026] (2) Based on the classification of weld point image features, this case also introduces multimodal information fusion (such as image, acoustic, electrical signal and thermal signal) to improve the detection capability of internal defects such as cold welds, reduce false alarm and false alarm rates, and at the same time support the coordinated adjustment of multiple welding parameters to optimize the welding process and adapt to weld head wear or working condition changes.

[0027] (3) In this case, a continuous state judgment mechanism was also introduced into the spot welding deep learning model to avoid frequent adjustments due to single misjudgment or instantaneous interference. The simulation verification was carried out in combination with the digital twin model in the welding point state ~ instruction parameter mapping rule library to ensure the safety and effectiveness of the parameter adjustment instruction. It also supports confidence judgment and only triggers the adjustment instruction when the model outputs a high confidence level.

[0028] (4) This case also sets up trend prediction and forward adjustment steps. By recording historical weld point image data, a weld point quality time series trend model is constructed, which can predict the wear trend of the weld head and make preventive parameter fine-tuning before the weld point abnormality occurs, so as to realize the optimization upgrade from passive reaction to active prediction.

[0029] (5) This case also sets up self-learning and rule optimization steps. The spot welding deep learning model continuously collects the adjusted weld point data to form a feedback dataset, thereby continuously incrementally training itself to optimize the recognition accuracy; at the same time, it adaptively adjusts the mapping relationship in the weld point state ~ instruction parameter mapping rule library to improve the self-adjustment effect. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the lead wire solder joint area of ​​the voice coil and FPC in a loudspeaker provided in a specific embodiment of the present invention;

[0031] Figure 2 This is a schematic diagram showing the different solder joint states of the lead wire solder joints of the voice coil and FPC in the loudspeaker provided in a specific embodiment of the present invention;

[0032] Figure 3 This is the early warning control display interface provided in a specific embodiment of the present invention after the solder joint status label is generated and before the parameter adjustment command is generated;

[0033] Figure 4 This is a schematic diagram of the structure of each module in a spot welding machine welding head self-adjustment system based on CCD-AI detection provided in a specific embodiment of the present invention;

[0034] Figure 5 This is a schematic diagram of the principle of a spot welding machine welding head self-adjustment system based on CCD-AI detection provided in a specific embodiment of the present invention. Detailed Implementation

[0035] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0036] To address the issue that spot welding (resistance welding) processes used in smaller electronic products (such as voice coils in speakers and lead soldering in FPCs) suffer from weak defect detection capabilities, particularly when the welding head is worn or under complex operating conditions (such as scratches or stains), resulting in low defect detection accuracy and high miss rates, and difficulty in timely and effective adjustment of spot welding machine parameters, this invention proposes a CCD-AI-based spot welding machine welding head self-adjustment method and system. The core of this invention is to observe the size of the weld marks on the spot welding machine through visual monitoring combined with a deep learning model, and then adjust the spot welding machine parameters (such as welding voltage) accordingly. This achieves automatic identification of the weld mark status and automatic adjustment of the spot welding machine parameters. Compared to the existing closed-loop adjustment process of image acquisition ~ feature extraction / defect identification ~ decision generation ~ parameter adjustment, this invention further optimizes the spot welding process for micro-electronic products. Based on the use of a spot welding deep learning model to classify weld mark statuses, a rule library mapping weld mark status to command parameters is established, allowing for more refined decision-making and compensation for welding head wear or changes in operating conditions based on changes in weld mark status.

[0037] Example 1: A self-adjusting method for spot welding machine head based on CCD-AI detection, comprising the following steps:

[0038] S00: Image Acquisition and Processing. The CCD-AI vision inspection module acquires images of the weld points after welding using a spot welding machine and processes these images to extract the weld point areas. The CCD-AI vision inspection module includes a camera, an industrial lens (configured with an image sensor, such as CCD or CMOS, to convert light signals into electrical signals), a light source, camera I / O cables, a network cable (or USB 3.0 cable), an industrial computer, and a monitor. This process uses the CCD-AI vision inspection module to acquire and form digital images, obtaining the original image (e.g., ...). Figure 1 (This is a schematic diagram of the solder joint area between the voice coil and the FPC in the loudspeaker). Further, the CogAffineTransformTool (which generates a new image through radial transformation) is used to extract the solder joint state area image. Thus, before proceeding to step S10, a spot welding deep learning model needs to be pre-built and trained so that it can output the state category label of the solder joint for subsequent decision-making, generating appropriate parameter adjustment instructions. The following details the spot welding deep learning model and its training process:

[0039] (1) Constructing a deep learning model for spot welding, including the following steps:

[0040] 1.1) Collect and extract at least 2000 feature images of the solder joint area, and classify and label them according to the solder joint status;

[0041] 1.2) Create a solder joint status learning project, such as Figure 2The image shows the annotation of solder joint status when creating a learning project. Each learning project includes: NG (Not Soldered) (unacceptable, no solder joint formed), NG (Poor Solder Joint) (unacceptable, small solder mark, solder joint size below the lower limit, insufficient strength), OK (Small Solder Mark) (acceptable, although the solder joint is slightly small, it is still within the process tolerance), NG (Large Solder Mark) (unacceptable, solder joint size exceeds the upper limit, may burn through or affect assembly), OK (Large Solder Mark) (acceptable, solder joint is slightly large but does not exceed the upper limit, acceptable), OK (Normal) (solder joint size, shape, and color are all near the standard center value). It should be noted that due to the different objects being soldered, the label may vary depending on the object being soldered, such as the voice coil in a speaker. Taking the lead soldering of FPC as an example, when the volume of the speaker is different, the corresponding voice coil and the normal solder joint state of the FPC lead are also different (including size, shape, color, etc.). Therefore, the "small solder mark" and "large solder mark" mentioned in this case are only used as classification labels. They have been manually labeled according to the working conditions in the early stage, without restricting them on a physical scale (it is also unreasonable to restrict them from a practical point of view). During the training process, the AI ​​model (spot welding deep learning model) has learned and internalized the image feature boundaries that distinguish these categories by using a large number of training images labeled as "small solder mark" or "large solder mark".

[0042] 1.3) Import the training images matched for each learning project, check the training images, and adjust the image size, aspect ratio, brightness, and contrast;

[0043] 1.4) Create a split, which determines how many labeled images to use for training, validation, and evaluation. The selected split ratio and the corresponding mapping of images to the dataset are saved in the project. Preferably, in this case, the image dataset is divided into the following proportions: 70% training set (images used for actual training), 15% validation set (images used to verify training progress), and 15% test set (images used to verify the quality of training results).

[0044] 1.5) Select the split project and set the selected training model algorithm. In this case, an AI deep learning model with strong feature extraction capabilities is needed, which is suitable for industrial and scientific research scenarios such as image classification, object detection, and semantic segmentation. Preferably, the spot welding deep learning model in this case is the Resnet-50 model.

[0045] (2) Training of the deep learning model for spot welding,

[0046] 2.1) Preliminary preparation: Set the image size to 200*200 pixels, configure GPU-based training, and use a deterministic algorithm; start training after setting the training parameters;

[0047] 2.2) Number of Epochs: The number of times a complete iteration is performed on the entire training data; Default value: 20

[0048] 2.3) Number of iterations: The number of times a single batch passes through the network, related to the number of epochs; Default value: Calcable;

[0049] 2.4) Batch size: The number of input images transferred to the device memory at one time; it should be noted that when using GPU training, generally speaking, the larger the batch size, the faster the evaluation. However, an excessively large batch size may have the opposite effect and significantly slow down the evaluation speed. The batch size must be adapted to the available memory.

[0050] 2.5) Learning rate: Step size; determines the weight of the gradient on the updated loss function arguments; default value: 0.001; a "change policy" can be defined: start with a relatively high value (e.g., 0.001) and then decrease it after a certain number of epochs (e.g., decrease from 0.00001 to 0.000001, ideally always decreasing this value when the loss no longer changes).

[0051] After training is complete, the model is evaluated and exported. First, other independent image sets can be selected instead of just the above image sets for evaluation, and the amount of evaluation data can be limited as needed, with an appropriate batch size set. Second, GPUs can be used to accelerate the evaluation, preferably using NVIDIA® TensorRT™ (GPU) or AI Accelerator. Then, the accuracy (number of correct predictions / total number of predictions) is calculated. Finally, after confirming the evaluation results, the model file and the accompanying dictionary / configuration file are exported.

[0052] Therefore, in the CCD~AI visual inspection module, a pre-trained AI model and its configuration file are integrated and called in the C#-developed CCD visual inspection software to perform the visual inspection task; further, step S10 is executed: state recognition and classification. The CCD~AI visual inspection module inputs the collected solder joint area image into the pre-trained spot welding deep learning model. The spot welding deep learning model outputs the state category label of the solder joint. The state category label includes normal, not welded, cold weld, and large solder mark. In summary, when the spot welding deep learning model recognizes the solder joint state as OK (small solder mark), OK (large solder mark), or OK (normal), it outputs the "normal" state category label; when the recognized solder joint state is NG (not welded), it outputs the "not welded" state category label; when the recognized solder joint state is NG (cold weld) or NG (small solder mark), it outputs the "cold weld" state category label; when the recognized solder joint state is... If the weld mark is large (NG), then the "large weld mark" status category label is output. In this step, the weld point status mentioned above is further classified for easier output. However, it should be emphasized that these status category labels in this case refer to the category labels output by the spot welding deep learning model, which correspond to the specific parameter adjustment instructions in the weld point status ~ instruction parameter mapping rule base in the next step S20. The specific discrimination criteria are determined by the sample set and model parameters used to train the model (i.e., they have been set during the early model training). The purpose is to trigger the corresponding self-adjustment action, rather than limiting the absolute geometric dimensions. On this basis, it should be noted that when the welding head has new unknown stains or the workpiece surface has a special coating, it may cause the visual image to show features that have never been seen before. Pure visual information is easy for the model to misjudge. Therefore, in order to improve the robustness of the spot welding deep learning model in classification and discrimination.

[0053] After the spot welding deep learning model accurately outputs the state category label of the weld point, it is necessary to generate a decision instruction to adjust the spot welding machine based on the weld point state, that is, to further execute step S20: adjustment decision generation. The CCD~AI vision detection module queries the preset weld point state~instruction parameter mapping rule library according to the state category label, generates a parameter adjustment instruction for the spot welding machine and sends it to the logic control module. The weld point state~instruction parameter mapping rule library is set with the adjustment value of the welding parameters of the spot welding machine corresponding to the state category label of the weld point. In step S10, by collecting single-modal and multi-modal information of weld point state image information, the category of weld point state can be accurately determined. However, individual misjudgments or isolated, non-trend defects may still occur due to unpredictable instantaneous interference. In particular, the wear of the welding head is gradually generated for the spot welding machine. That is, when the spot welding machine needs to be adjusted using parameter adjustment instructions, the wear of the welding head of the spot welding machine will continue and become more and more frequent, leading to a state of poor welding or smaller weld points, or abnormal states will occur multiple times. Therefore, before generating the parameter adjustment instruction for the spot welding machine, the following steps are also included:

[0054] S21: Continuous state judgment, statistical analysis of the state identification results of N solder joints produced in continuous production;

[0055] S22: Trigger condition determination: The parameter adjustment command for the spot welding machine will be generated and an early warning will be triggered only when M consecutive abnormal weld points of the same type occur, or when the frequency of abnormal occurrence exceeds the preset threshold. M and N are preset positive integers, and M≤N.

[0056] By employing steps S21 and S22, the stability of the parameter adjustment commands generated based on the solder joint status can be greatly improved. This prevents sudden changes or the status of a single solder joint from affecting the continuous welding process. Figure 3 The image shows its early warning control display interface, which places greater emphasis on the overall stability of the welding process. Of course, if greater emphasis is placed on the sensitivity of the welding process, the values ​​of M and N can be reduced depending on the situation, which will not be elaborated here.

[0057] Preferably, the weld joint status ~ instruction parameter mapping rule library is a quantitative mapping library, and the parameter adjustment instruction, taking welding voltage as an example, includes at least the following instructions:

[0058] The "not welded" status label is mapped to an instruction indicating that the spot welding machine's welding head is damaged.

[0059] The cold solder joint status label is mapped to an increase in soldering voltage. The instructions also include a small solder stamp status label mapped to an increase in soldering voltage. The instructions, in which ;

[0060] Large solder mark status label mapped to reduced soldering voltage The instructions;

[0061] The normal status label is further divided into an OK status label, an OK solder mark small label, and an OK solder mark large label. The OK status label is mapped to the spot welding machine voltage normal command, and the OK solder mark small label is mapped to the increase of welding voltage. The instruction that the OK soldering mark is mapped to an increased soldering voltage The instructions; based on this, taking a speaker with a model size of 17×12mm as an example, the relationship between the solder joint status and the spot welding machine voltage is as follows:

[0062] No welding → Damaged spot welding head;

[0063] Cold weld → The spot welding machine voltage is extremely low; increase the voltage by 0.5V.

[0064] Small weld mark → Spot welding machine voltage is too low, increase voltage by 0.3V.

[0065] Large solder joint → Spot welding machine voltage is extremely high, reduce voltage by 1.0V.

[0066] OK status → Spot welding machine voltage is normal

[0067] OK solder mark small → Spot welding machine voltage is too low, increase voltage by 0.3V;

[0068] OK solder mark is large → Spot welding machine voltage is too high, reduce voltage by 0.8V;

[0069] Based on the above instructions, this case can further construct a relationship between the feature values ​​of the weld point image and the voltage adjustment of the spot welding machine:

[0070]

[0071] in, For the final calculated welding voltage adjustment (unit: V), when When it indicates that the voltage needs to be reduced, when This indicates that the voltage needs to be increased; The reference adjustment strength coefficient (unit: V) represents the amount of basic voltage adjustment that needs to be applied when the relative deviation of the weld joint size is 100%. Its initial value can be determined by operating conditions or experimental data. The weld mark feature size is actually measured from the current weld point image by the spot welding deep learning model. It can be the weld point area size, the number of pixels in the equivalent area, etc. The characteristic dimension of the target solder joint (i.e., the value of the solder joint in the OK state) is a reference value preset according to the product process standard; These are dynamic parameters that vary with the number of welding passes, used to optimize the welding voltage adjustment. ; To compensate for constants.

[0072] The previous description proceeded from image acquisition of the weld joint area to weld joint state type identification. Then, based on the weld joint state type identification label of the spot welding deep learning model, the parameter adjustment command was generated. Before the parameter adjustment command was sent to the spot welding machine for execution, to further improve safety and reliability (constraints were already added during weld joint state type identification), and to compensate for potential uncertainties in the spot welding deep learning model, in step S20, the weld joint state-command parameter mapping rule base also includes a digital twin model. After generating the parameter adjustment command for the spot welding machine, the parameter adjustment command and the current operating condition are input into the digital twin model of the spot welding process for simulation. If the simulation result predicts that the weld joint quality meets the requirements and there is no equipment risk, then step S30 is confirmed to continue; otherwise, the parameter adjustment command is corrected or canceled, and an early warning is issued. This achieves dual guarantee decision-making through the spot welding deep learning model and the digital twin model in the weld joint state-command parameter mapping rule base, enabling the executable parameter adjustment commands to more effectively address problems caused by welding head wear or complex operating condition changes.

[0073] After obtaining the executable parameter adjustment instructions through the above steps, step S30 is performed: instruction execution and adjustment. The logic control module sends the parameter adjustment instructions to the controller of the spot welding machine, driving the spot welding machine to automatically adjust the welding parameters to compensate for wear of the welding head or changes in working conditions.

[0074] Example 2: In Example 1, parameter adjustment commands for the spot welding machine were mainly generated based on the image information of the weld joint area acquired by the CCD~AI vision detection module. This involved single-modal discrimination using the CCD camera to visually acquire the image information of the weld joint area. To overcome the limitations of single-image feature parameter discrimination and enable the spot welding deep learning model to more accurately determine the weld joint state category label, this example introduces multi-modal discrimination. The spot welding deep learning model identifies the weld joint state based on the fusion of image information acquired by the CCD~AI vision detection module and non-image information from at least one auxiliary sensor. The non-image information includes welding... The acoustic signal (which can characterize metal melting, spatter, and the generation of microcracks, and is extremely sensitive to defects such as "cold welds" due to poor internal bonding), electrical signal (current / voltage waveforms can accurately reflect whether the input energy is sufficient and stable, and whether the contact resistance is abnormal), or thermal signal (infrared thermometry can reflect whether the heat distribution is uniform, and whether there is local undercooling or overheating) can be used to determine the state category label of the weld point and output it. This allows the spot welding deep learning model to not only use single-modal discrimination based on visual image information acquired by a CCD camera, but also to combine other non-image information for multi-modal comprehensive discrimination. For example, the sensitivity of acoustic and electrical signals to internal defects such as cold welds is much higher. From a visual perspective, determining whether a weld bead is too large or too small, combined with energy signals, can differentiate between "inappropriate parameters" and "material thickness fluctuations," making the judgment more accurate and significantly improving defect identification accuracy while reducing false alarms / missed detections. Furthermore, CCD cameras capture the static appearance of the weld after completion. Their perceptual and analytical capabilities are insufficient when distinguishing between defects with similar appearances but different causes (such as "cold weld" and "insufficient heat due to slight weld head contamination," both of which may appear as small weld beads in the image). By introducing multimodal information, deep learning models for spot welding can learn the complete cause of the defect, thus achieving a more fundamental and accurate understanding of the weld bead state. Accurate judgments make the generation of parameter adjustment instructions for the spot welding machine in subsequent steps more precise. Furthermore, the spot welding deep learning model outputs a weld point state category label and its corresponding comprehensive confidence level, i.e., the degree of certainty of the current judgment made by the spot welding deep learning model on the output weld point state category label. When the comprehensive confidence level is higher than a preset threshold, the parameter adjustment instruction is generated based on the state category label in step S20. This sets a safety redundancy for the decision of the spot welding deep learning model to output the weld point state category label, so that the parameter adjustment instruction can only be generated when the weld point state category has high certainty, ensuring the accuracy of the subsequent self-adjustment process.

[0075] Based on the above description, and by combining single-modal acquisition of weld point state image information with non-image information for multimodal weld point state classification, to address the limitations of single image feature parameter discrimination and corresponding parameter adjustment instructions, in step S20, according to the identified weld point state category label and the image features of the weld point area, the optimal collaborative adjustment amount of each parameter is obtained through a multi-parameter optimized spot welding deep learning model. The parameter adjustment instructions include instructions for collaborative adjustment of at least two parameters among welding voltage, electrode pressure, and welding time. Thus, the original rule mapping based on single-modal discrimination (i.e., detection-feedback control) is transformed into a collaborative optimization method based on multimodal discrimination. That is, the spot welding deep learning model will comprehensively judge the acquired multimodal information, observe the image features of the weld point area, and understand the physical mechanism of weld point state generation in conjunction with other non-image information, and then generate the optimal instruction for parameter adjustment to solve the compensation problem for welding head wear or special working conditions. Based on this, the weld point quality control of the spot welding process can also be further autonomously optimized and continuously evolved.

[0076] Example 3: In Examples 1 and 2, the CCD-AI vision detection module acquires images of the weld area and generates corresponding parameter adjustment commands for the spot welding machine. These are targeted responses after a certain state of the weld has occurred. To effectively adjust before undesirable states (i.e., no welding, cold welding, large weld marks) occur, step S40 is added to Examples 1 and 2: trend prediction and forward-looking adjustment. This involves recording and analyzing the feature data of historical weld images to construct a time-series trend of weld quality changes. Based on this trend, a prediction model is used to determine the wear trend of the welding head or the stability of the process. When the prediction result indicates that the risk of quality degradation exceeds a preset threshold, a forward-looking preventive parameter fine-tuning command is generated and executed before the weld state is identified as abnormal. This allows the spot welding machine's welding head self-adjustment to not only respond passively but also proactively predict and adjust, i.e., issue parameter fine-tuning commands. The entire self-adjustment method can not only address the problem of weld quality degradation caused by welding head wear but also ensure that the spot welding process remains in optimal condition for a long time, predicting and avoiding undesirable states.

[0077] Example 4: Through steps S00-S30, the spot welding machine can adjust according to the detection and parameter adjustment instructions of the CCD-AI vision inspection module to address and compensate for problems caused by welding head wear and other working conditions. Step S40 can further perform forward-looking predictive adjustment. In these steps, as the usage time and welding times increase, more and more data are accumulated. Based on this, step S50 is also included: self-learning and rule optimization, continuously collecting weld point images and their state data after executing adjustment instructions to form a feedback dataset; based on the feedback dataset, the spot welding deep learning model is incrementally trained to optimize the recognition accuracy, and / or the mapping relationship in the weld point state-instruction parameter mapping rule base is adaptively adjusted to optimize the adjustment effect. In this way, the entire self-adjustment process can be self-optimized, and its recognition accuracy and adjustment effect will become better and better.

[0078] Example 5: A self-adjusting system for a spot welding machine head based on CCD-AI detection, used to implement the self-adjusting method for a spot welding machine head based on CCD-AI detection as described above, such as... Figure 4 and Figure 5 The following modules are shown:

[0079] The CCD-AI vision inspection module includes an industrial computer and an inspection camera for acquiring weld point images. The industrial computer has a built-in AI processing module for running a spot welding deep learning model and a weld point state-instruction parameter mapping rule library to perform state recognition and generate parameter adjustment instructions.

[0080] The logic control module is used to receive the parameter adjustment instructions and convert them into control signals;

[0081] The spot welding machine module is controlled by the control signal to adjust its welding parameters.

[0082] The logic control module is a PLC control system. The PLC control system communicates with the CCD~AI vision inspection module and the spot welding machine module through an industrial communication protocol. It is used to parse the parameter adjustment instructions and generate specific control pulse sequences. Preferably, after receiving the signal, the PLC sends parameter adjustment instructions according to Modbus. The spot welding machine receives the parameter adjustment instructions through Modbus or Ethernet protocol and adjusts the welding parameters (such as welding voltage).

[0083] This invention has been described through preferred embodiments. Those skilled in the art will understand that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. This invention is not limited to the specific embodiments disclosed herein; other embodiments falling within the scope of the claims are also within the protection scope of this invention.

Claims

1. A self-adjusting method for the welding head of a spot welding machine based on CCD-AI detection, characterized in that, Includes the following steps: S00: Image acquisition and processing. The CCD~AI vision inspection module acquires images of the weld points after welding by the spot welding machine and processes the images to extract the weld point area. S10: Status recognition and classification. The CCD~AI vision detection module inputs the collected solder joint area image into the pre-trained spot welding deep learning model. The spot welding deep learning model outputs the status category label of the solder joint. The status category label includes normal, unwelded, cold solder joint and large solder mark. S20: Adjustment Decision Generation. The CCD~AI vision detection module queries the preset weld point state~instruction parameter mapping rule library according to the state category label, generates parameter adjustment instructions for the spot welding machine, and sends them to the logic control module. The weld point state~instruction parameter mapping rule library is set with the welding parameter adjustment values ​​of the spot welding machine corresponding to the state category label of the weld point. Before generating the parameter adjustment instructions for the spot welding machine, the optimal coordinated adjustment amount of each parameter is obtained through a multi-parameter optimized spot welding deep learning model based on the identified weld point state category label and the image features of the weld point area. The parameter adjustment instructions include instructions for coordinated adjustment of at least two parameters among welding voltage, electrode pressure, and welding time. Also includes: S21: Continuous state judgment, statistical analysis of the state identification results of N solder joints produced in continuous production; S22: Trigger condition determination: The parameter adjustment command for the spot welding machine will be generated and an early warning will be triggered only when M consecutive abnormal weld points of the same type occur, or when the frequency of abnormal occurrence exceeds the preset threshold. M and N are preset positive integers, and M≤N. Trend prediction and forward-looking adjustment: Record and analyze the feature data of historical solder joint images to construct the time-series trend of solder joint quality changes; Based on the trend, determine the wear trend of the solder head or the stability of the process through a prediction model; When the prediction result indicates that the risk of quality degradation exceeds a preset threshold, generate and execute a forward-looking preventive parameter fine-tuning instruction before the solder joint status is identified as abnormal. The solder joint status ~ instruction parameter mapping rule base is a quantitative mapping library, which includes at least the following instructions: The "not welded" status label is mapped to an instruction indicating that the spot welding machine's welding head is damaged. The cold solder joint status label is mapped to an increase in soldering voltage. The instructions also include a small solder stamp status label mapped to an increase in soldering voltage. The instructions, in which ; Large solder mark status label mapped to reduced soldering voltage The instructions; The normal status label is further divided into an OK status label, an OK solder mark small label, and an OK solder mark large label. The OK status label is mapped to the spot welding machine voltage normal command, and the OK solder mark small label is mapped to the increase of welding voltage. The instruction that the OK soldering mark is mapped to an increased soldering voltage The instructions; The relationship between the feature values ​​of the weld spot image and the voltage adjustment of the spot welding machine is as follows: in, For the final calculated welding voltage adjustment, when When it indicates that the voltage needs to be reduced, when This indicates that the voltage needs to be increased; The reference adjustment strength coefficient represents the amount of basic voltage adjustment that needs to be applied when the relative deviation of the weld joint size is 100%. Its initial value is determined by the operating conditions. The weld mark feature size is the actual measurement obtained from the current weld point image by the spot welding deep learning model; The characteristic dimensions of the target solder joint are reference values ​​pre-set according to the product process standards; These are dynamic parameters that vary with the number of welding passes, used to optimize the welding voltage adjustment. ; To compensate for constants; The weld point status ~ instruction parameter mapping rule base is also equipped with a digital twin model. After generating parameter adjustment instructions for the spot welding machine, the parameter adjustment instructions and the current working condition are input into the digital twin model of the spot welding process for simulation. If the simulation result predicts that the weld point quality meets the requirements and there is no equipment risk, then it is confirmed to continue executing step S30; otherwise, the parameter adjustment instruction is corrected or canceled and an early warning is issued. S30: Instruction execution and adjustment. The logic control module sends the parameter adjustment instruction to the controller of the spot welding machine, driving the spot welding machine to automatically adjust the welding parameters to compensate for wear of the welding head or changes in working conditions.

2. The self-adjusting method for a spot welding machine head based on CCD-AI detection according to claim 1, characterized in that, In step S10, the spot welding deep learning model identifies the weld point status based on the image information collected by the CCD~AI vision detection module and non-image information from at least one auxiliary sensor. The non-image information includes acoustic signals, electrical signals, or thermal signals during welding. The spot welding deep learning model outputs a weld point status category label and a corresponding comprehensive confidence level. Only when the comprehensive confidence level is higher than a preset threshold is a parameter adjustment instruction generated based on the status category label in step S20.

3. The self-adjusting method for a spot welding machine head based on CCD-AI detection according to claim 1, characterized in that, It also includes step S50: self-learning and rule optimization, continuously collecting weld point images and their state data after executing adjustment instructions to form a feedback dataset; based on the feedback dataset, incrementally training the spot welding deep learning model to optimize recognition accuracy, and / or adaptively adjusting the mapping relationship in the weld point state ~ instruction parameter mapping rule base to optimize the adjustment effect.

4. A self-adjusting system for a spot welding machine head based on CCD-AI detection, used to implement the self-adjusting method for a spot welding machine head based on CCD-AI detection as described in any one of claims 1 to 3, characterized in that, Includes the following modules: The CCD-AI vision inspection module includes an industrial computer and an inspection camera for acquiring weld point images. The industrial computer has a built-in AI processing module for running a spot welding deep learning model and a weld point state-instruction parameter mapping rule library to perform state recognition and generate parameter adjustment instructions. The logic control module is used to receive the parameter adjustment command and convert it into a control signal; The spot welding machine module is controlled by the control signal to adjust its welding parameters.

5. The self-adjusting system for a spot welding machine head based on CCD-AI detection according to claim 4, characterized in that, The logic control module is a PLC control system. The PLC control system communicates with the CCD~AI vision inspection module and the spot welding machine module through an industrial communication protocol, and is used to parse the parameter adjustment instructions and generate specific control pulse sequences.

Images