Sheet metal part welding quality detection device and system based on artificial intelligence and automatic control technology

Abstract

The invention relates to the technical field of sheet metal part welding quality detection, and discloses a sheet metal part welding quality detection device based on artificial intelligence and automatic control technology, which comprises an automatic loading and unloading mechanism, a conveying mechanism, a high-resolution camera and a server, the invention further discloses a sheet metal part welding quality detection system based on the artificial intelligence and the automatic control technology, the sheet metal part welding quality detection system comprises a hardware unit and a software unit, and the sheet metal part welding quality detection process is disclosed. The sheet metal part welding quality is efficiently and accurately detected, the production efficiency and the product quality are greatly improved, the advanced automatic control technology and the deep learning algorithm are combined, the quality management level of the manufacturing industry is remarkably improved, and coordinated operation of the whole system is ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of sheet metal welding quality inspection, and specifically to a sheet metal welding quality inspection device and system based on artificial intelligence and automation control technology. Background Art

[0002] The main purpose of sheet metal welding quality inspection is to ensure the firmness, uniformity and integrity of the weld, prevent defects such as false welding, cracks, incomplete penetration, burn-through, notch, and undercut, so as to ensure the overall quality and reliability of the sheet metal.

[0003] After retrieval, a patent with the application number CN202020550067.3 discloses a temporary inspection device for the welding of sheet metal parts used in automobile assembly, including a support frame. A handle is sleeved in the middle of the support frame. Second connection blocks are fastened at both ends of the surface of the support frame. Guide cylinders are installed at both ends of the support frame. A rotating block is connected between the guide cylinder and the support frame. A first connection block is welded on the side surface of the rotating block. A first spiral spring is welded between the first connection block and the second connection block. A frame is welded on the outer side surface of the guide cylinder, and a motor is fixedly installed at the outer end of the frame. This utility model is convenient to operate. By using the drive of the motor, the knocking plate can be used to repeatedly knock the automobile sheet metal parts, and the welding quality of the welding part can be detected by vibration, which is more intuitive and fast.

[0004] Currently, the following several methods are mainly used for sheet metal welding quality inspection:

[0005] 1. Visual inspection: This is the most traditional inspection method, which relies on the operator to observe the color, shape, continuity and other characteristics of the welding part with the naked eye to judge whether the welding quality is qualified. This method is simple and easy to implement, but there are problems such as strong subjectivity, easy omission and misjudgment.

[0006] 2. Ultrasonic testing: High-frequency sound waves are emitted by an ultrasonic probe. When the sound waves propagate inside the metal and encounter defects, reflections will occur, so as to detect the defects inside the weld. This method can detect internal cracks, pores and other defects, but the equipment cost is relatively high, and professional technical personnel are required to operate and interpret the results.

[0007] 3. Radiographic testing: X-rays or γ-rays are used to penetrate the weld, and the defects inside the weld are displayed through film imaging. This method has high detection accuracy, but there are radiation safety problems, and the detection speed is slow, which is not suitable for large-scale production environments.

[0008] 4. Magnetic particle testing: It is applicable to ferromagnetic materials. By applying magnetic powder on the surface of the weld, the defects on the surface and near the surface are detected by using the magnetic field change. This method is simple to operate, but is limited to ferromagnetic materials and is difficult to detect internal defects.

[0009] 5. Penetrant testing: The penetrant liquid seeps into the tiny defects on the surface of the weld, and then the developer is used to display the defect positions. This method is applicable to various materials, but it can only detect surface-opening defects, and the detection accuracy is greatly affected by the operator's experience.

[0010] Although the above methods can meet the requirements of sheet metal welding quality inspection to a certain extent, there are still the following problems:

[0011] 1. Strong subjectivity: Methods such as visual inspection and penetrant testing highly rely on the experience and skills of operators. There may be significant differences in the inspection results among different operators, resulting in low stability and reliability of the inspection results.

[0012] 2. Low inspection efficiency: Although methods such as ultrasonic testing and radiographic testing have high inspection accuracy, their inspection speed is slow, which is not suitable for large-scale production environments. Especially in assembly line operations, it will seriously affect production efficiency.

[0013] 3. High cost: Methods such as radiographic testing and ultrasonic testing require expensive equipment and professional technical personnel, increasing the operating costs of enterprises. Especially for small and medium-sized enterprises, the economic burden is relatively heavy.

[0014] 4. Safety issues: Radiographic testing has radiation safety risks and requires strict protective measures, increasing the operation complexity and management difficulty.

[0015] 5. Limited inspection range: Methods such as magnetic particle testing and penetrant testing can only detect surface or near-surface defects and cannot comprehensively evaluate the overall quality of the weld.

[0016] 6. Low degree of automation: Most of the existing inspection methods rely on manual operation, with a low degree of automation, making it difficult to achieve intelligent and unmanned production and unable to meet the development needs of modern manufacturing.

[0017] In summary, the existing sheet metal welding quality inspection methods have many deficiencies in terms of inspection accuracy, efficiency, cost, safety, and degree of automation.

[0018] Therefore, we need to propose a sheet metal welding quality inspection device and system based on artificial intelligence and automation control technology, upgrading from traditional manual visual inspection to artificial intelligence automatic inspection to improve inspection accuracy, precision, and inspection efficiency. Summary of the Invention

[0019] The object of the present invention is to provide a sheet metal welding quality detection device and system based on artificial intelligence and automatic control technology. Through the coordinated work of an automatic loading and unloading mechanism, a conveying mechanism, a high-resolution camera, a server, and auxiliary equipment, the efficient and accurate detection of sheet metal welding quality is realized, greatly improving the production efficiency and product quality. Combining advanced automatic control technology and deep learning algorithms significantly enhances the quality management level of the manufacturing industry and ensures the coordinated operation of the entire system to solve the problems raised in the above background technology.

[0020] To achieve the above object, the present invention provides the following technical solutions: A sheet metal welding quality detection device based on artificial intelligence and automatic control technology, comprising:

[0021] An automatic loading and unloading mechanism: responsible for automatically transporting the sheet metal parts to be detected to the detection area and removing the workpieces from the detection area after the detection is completed;

[0022] A conveying mechanism: used to transfer the sheet metal parts from the automatic loading and unloading mechanism to a transparent glass-like rotating detection table;

[0023] A high-resolution camera: used to take multi-angle pictures of the sheet metal parts on the rotating detection table to obtain high-quality sheet metal part image data;

[0024] A server: receives and processes the image data from the high-resolution camera and performs quality detection through a pre-trained neural network model;

[0025] The automatic loading and unloading mechanism and the rotating detection table are respectively seamlessly connected to both ends of the conveying mechanism. The high-resolution camera is electrically connected to the server. A plurality of high-resolution cameras are arranged around the rotating detection table, and the high-resolution camera is set to a camera with high resolution, fast frame rate, and good color restoration.

[0026] Preferably, the automatic loading and unloading mechanism includes a grasping robotic arm driven by a high-precision servo motor. The load range of the grasping robotic arm is 50 - 200KG, the repeat positioning accuracy of the grasping robotic arm is ±0.05mm, and the grasping robotic arm is integrated with the control system through a communication interface.

[0027] Preferably, the conveying mechanism includes a belt conveyor. The width of the belt on the belt conveyor is 800mm - 1500mm, and the feeding speed of the belt conveyor can be adjusted between 0.5m / s - 3m / s through a frequency converter;

[0028] There are six high-resolution cameras, including two left detection cameras, two right detection cameras, one top detection camera, and one back detection camera. The resolution of the high-resolution cameras is ≥20MP, the frame rate is ≥90fps, and they are equipped with fixed-focus or zoom lenses with distortion <1%.

[0029] Preferably, the server includes:

[0030] Processor: A multi-core high-performance CPU with a main frequency of the CPU ≥3.2GHz;

[0031] Graphics card: Containing 32GB HBM2 video memory to accelerate model inference;

[0032] Memory: Including an SSD of no less than 1TB for the operating system and temporary storage, and an NVMe SSD for array storage of the training data set.

[0033] Based on the above-described sheet metal welding quality detection device based on artificial intelligence and automation control technology, the present invention also provides a sheet metal welding quality detection system based on artificial intelligence and automation control technology, including:

[0034] Hardware unit, including

[0035] An automatic loading and unloading module that automatically transfers the sheet metal to be detected from the production line to the detection area and removes it from the detection area after the sheet metal detection is completed;

[0036] A conveying module that transfers the sheet metal from the automatic loading and unloading module to the rotating detection table;

[0037] An image acquisition module that takes multi-angle photos of the sheet metal on the rotating detection table;

[0038] A server that receives and processes the image data obtained by the image acquisition module and detects the image data;

[0039] And an auxiliary module, which includes a lighting lamp, a protective cover, and an audible and visual alarm;

[0040] The automatic loading and unloading module, the conveying mechanism, and the rotating detection table are arranged in sequence. The image acquisition module is installed on the rotating detection table, and the image acquisition module is electrically connected to the server;

[0041] Software unit, including

[0042] A data preprocessing module that cleans and labels and trains the obtained sheet metal image data;

[0043] A neural network model that detects the quality of the preprocessed sheet metal image;

[0044] A training module for training a neural network model;

[0045] A test and validation module for evaluating a neural network model.

[0046] Preferably, the workflow of the data preprocessing module includes:

[0047] Data collection: Collect sheet metal part images from different welding processes, materials, and environmental conditions;

[0048] Data cleaning: Include removing outliers, normalizing and standardizing, and resizing the sheet metal part images;

[0049] Data augmentation: Apply random rotation, flipping, scaling, and cropping transformations to expand the dataset and improve the generalization ability of the neural network model;

[0050] Annotation: Workers check each sheet metal part image, mark those with qualified welding quality as qualified, those with unqualified welding quality as unqualified, and annotate the reasons for unqualified.

[0051] Preferably, the neural network model includes:

[0052] Input layer: Responsible for receiving image data as the starting point of the neural network model, and the input dimension depends on the size of the image;

[0053] Hidden layer: Includes convolutional layer, pooling layer, fully connected layer, Dropout layer;

[0054] Output layer: Apply the Softmax function to convert it into a probability distribution, representing the prediction probability of each class.

[0055] Preferably, when the training module trains the neural network model, the contents used include:

[0056] Supervised learning algorithm: Given a labeled dataset, let the neural network model learn the mapping relationship from input to output;

[0057] Loss function: Use cross-entropy loss to measure the difference between the predicted distribution and the true distribution;

[0058] Optimizer: Use good convergence speed and adaptive learning rate characteristics to minimize the loss function;

[0059] Regularization techniques: Include L2 regularization, early stopping, used to avoid overfitting and improve the performance of the neural network model.

[0060] Preferably, the process for the test and validation module to evaluate the neural network model is as follows:

[0061] Split the dataset: Divide the original dataset into a training set, a validation set, and a test set to ensure there are enough samples for evaluating the model's generalization ability;

[0062] Evaluation metrics: Use four important metrics, namely accuracy, precision, recall, and F1-score, to measure the performance of the classification model, and use a confusion matrix to analyze misclassifications between different categories, paying attention to the detection accuracy of non-conforming categories;

[0063] Continuous improvement: Adjust the hyperparameters according to the performance on the validation set, retrain the model, and finally evaluate the actual effectiveness of the model using the test set.

[0064] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0065] Through the coordinated operation of the automatic loading and unloading mechanism, the conveying mechanism, the high-resolution camera, the server, and the auxiliary equipment, the present invention realizes the efficient and accurate detection of the welding quality of sheet metal parts, greatly improves the production efficiency and product quality, and combines advanced automation control technology and deep learning algorithms to significantly improve the quality management level of the manufacturing industry and ensure the coordinated operation of the entire system. Brief Description of the Drawings

[0066] Figure 1 is a schematic structural diagram of the present invention;

[0067] Figure 2 is a system block diagram of the present invention;

[0068] Figure 3 is a flowchart of the welding quality detection of sheet metal parts of the present invention. Detailed Description of the Invention

[0069] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0070] Please refer to Figures 1-3 , the present invention provides a technical solution: a sheet metal part welding quality detection device based on artificial intelligence and automation control technology, including:

[0071] Automatic loading and unloading mechanism: responsible for automatically transporting the sheet metal parts to be detected to the detection area and removing the workpieces from the detection area after the detection is completed;

[0072] The automatic loading and unloading mechanism includes a grasping robotic arm driven by a high-precision servo motor. The load range of the grasping robotic arm is 50 - 200KG, and the repeat positioning accuracy of the grasping robotic arm is ±0.05mm, ensuring the consistency and accuracy of each operation. The grasping robotic arm is integrated with the control system through a communication interface, and the communication interface supports industrial network protocols such as Ethernet / IP and Profinet.

[0073] The automatic loading and unloading mechanism usually adopts automated equipment such as robotic arms or conveyor belts, and realizes precise positioning and grasping through sensors and control systems. When the sheet metal part reaches the specified position, the robotic arm places it on the conveying mechanism to complete the loading process. After the detection is completed, the robotic arm grabs the sheet metal part again and moves it out of the detection area to complete the unloading process.

[0074] Conveying mechanism: used to transfer the sheet metal part from the automatic loading and unloading mechanism to the transparent glass-like rotating inspection table;

[0075] The conveying mechanism includes a belt conveyor. The width of the belt on the belt conveyor is 800mm - 1500mm. The length of the belt conveyor is designed to accommodate the longest workpiece and leave enough buffer space. The feeding speed of the belt conveyor can be adjusted between 0.5m / s - 3m / s through a frequency converter, and the belt conveyor is driven by an AC motor, and the power of the AC motor is 0.75KW - 4KW;

[0076] The conveying mechanism usually adopts a belt conveyor and realizes continuous or intermittent transmission through motor drive. The speed and direction of the belt conveyor can be adjusted through the control system to adapt to sheet metal parts of different sizes and shapes. The end of the belt conveyor is seamlessly docked with the transparent glass rotating inspection table to ensure the smooth transition of the sheet metal part.

[0077] High-resolution camera: used to take multi-angle photos of the sheet metal part on the rotating inspection table to obtain high-quality sheet metal part image data; and the high-resolution camera is set as a camera with high resolution, fast frame rate and good color restoration.

[0078] The high-resolution camera can select the Basler ace series or Sony Alpha series, providing multiple sensor options (CMOS / CCD).

[0079] There are six high-resolution cameras. The high-resolution cameras include two left detection cameras, two right detection cameras, one top detection camera, and one back detection camera. The resolution of the high-resolution camera ≥20MP to ensure detailed capture. The frame rate of the high-resolution camera ≥90fps. The high-resolution camera is equipped with a fixed-focus or zoom lens, and the field of view angle covers the entire surface of the sheet metal part, and the distortion of the lens <1%.

[0080] Through the distribution of six high-resolution cameras, it is ensured that all surfaces of the sheet metal parts can be clearly captured. The high-resolution cameras are connected to the server through high-speed interfaces, and the captured image data is transmitted in real time. The resolution and frame rate of the high-resolution cameras can be adjusted according to the detection requirements to ensure the clarity and integrity of the images.

[0081] Server: Receive and process the image data from the high-resolution cameras, and perform quality detection through a pre-trained neural network model;

[0082] The server includes:

[0083] Processor: Adopt a multi-core high-performance CPU, and the main frequency of the CPU ≥ 3.2 GHz;

[0084] Graphics card: Include 32GB HBM2 video memory to accelerate model inference;

[0085] Memory: Include an SSD of no less than 1TB for the operating system and temporary storage, and an NVMe SSD for array storage of the training data set.

[0086] And the server is connected through Gigabit Ethernet or a faster optical fiber to ensure real-time data stream transmission. The server is also equipped with a cooling device (liquid cooling or air cooling) to keep the system running stably.

[0087] The automatic loading and unloading mechanism and the rotary detection table are respectively seamlessly connected to both ends of the conveying mechanism. The high-resolution cameras are electrically connected to the server, and a plurality of high-resolution cameras are arranged around the rotary detection table.

[0088] The server is equipped with a high-performance processor and a large-capacity storage device, which can quickly process a large amount of image data. After the image data is transmitted to the server through a high-speed network, the server runs a neural network model built based on TensorFlow for analysis. The model outputs 10 categories: qualified and various unqualified (such as cracks, pores, incomplete penetration, etc.). The detection results are displayed on the display screen, and the alarm system is triggered when unqualified products are found.

[0089] Through the cooperation and coordinated work of the above devices, the efficient and accurate detection of the welding quality of sheet metal parts is realized, greatly improving the production efficiency and product quality.

[0090] Based on the above-described sheet metal part welding quality detection device based on artificial intelligence and automation control technology, the present invention also provides a sheet metal part welding quality detection system based on artificial intelligence and automation control technology, including:

[0091] Hardware unit, including

[0092] An automatic loading and unloading module that automatically transfers the sheet metal parts to be detected from the production line to the detection area and removes them from the detection area after the detection of the sheet metal parts is completed;

[0093] A conveying module that transfers the sheet metal parts from the automatic loading and unloading module to the rotating detection table;

[0094] An image acquisition module that takes multi-angle photos of the sheet metal parts on the rotating detection table;

[0095] A server that receives and processes the image data acquired by the image acquisition module and detects the image data;

[0096] And an auxiliary module, the auxiliary module includes a lighting lamp, a protective cover, and an audible and visual alarm;

[0097] Among them, in order to improve the image quality, the lighting lamp needs to be equipped with a uniform and brightness-adjustable LED ring lamp or other professional light sources; the waterproof and dustproof level of the protective cover reaches above IP65, and it is installed around the high-resolution camera to prevent the high-resolution camera from being contaminated by dust and oil in the production environment. When it is detected that the welding of the sheet metal parts is unqualified, the audible and visual alarm operates to notify the operator in time.

[0098] The automatic loading and unloading module, the conveying mechanism, and the rotating detection table are arranged in sequence. The image acquisition module is installed on the rotating detection table, and the image acquisition module is electrically connected to the server;

[0099] The devices in the hardware unit of the present invention comply with their respective technical standards and are carefully matched to ensure the coordinated operation of the entire system, realizing the efficient and accurate detection of the welding quality of sheet metal parts.

[0100] A software unit, including

[0101] A data preprocessing module that cleans and labels and trains the acquired sheet metal part image data;

[0102] The working process of the data preprocessing module includes:

[0103] Data collection: Collect sheet metal part images under different welding processes, materials, and environmental conditions to ensure the diversity and representativeness of the data set, and segment the collected images, combine POEN CV to identify the circular contour of the solder joints, segment the images according to the solder joints and number them;

[0104] Data cleaning: including removing outliers, standardizing and normalizing processing, and size adjustment of the sheet metal part images;

[0105] Among them, removing outliers: identifying and excluding obviously incorrect or irrelevant samples (such as blurred, severely tilted, etc.) to reduce the negative impact on model training.

[0106] Standardization / Normalization: Adjust the image pixel values to the range of 0 - 1 for subsequent processing and to maintain consistency across different batches.

[0107] Resizing: Resize all images to a fixed size (256x256 pixels) to facilitate input into the neural network.

[0108] Data Augmentation: Apply random rotation, flipping, scaling, and cropping transformations to expand the dataset and improve the generalization ability of the neural network model.

[0109] Annotation: Workers check each sheet metal part image, mark those with qualified welding quality as qualified, those with unqualified welding quality as unqualified, and annotate the reasons for unqualified.

[0110] A neural network model for quality inspection of pre - processed sheet metal part images;

[0111] The neural network model includes:

[0112] Input Layer: Responsible for receiving image data as the starting point of the neural network model. The input dimension depends on the size of the image.

[0113] Hidden Layers: Include convolutional layers, pooling layers, fully - connected layers, and Dropout layers.

[0114] Among them, the convolutional layer: Uses 128 convolutional operations to extract features. Each layer contains a certain number of filters (such as 32, 64, 128), and the activation function usually selects ReLU.

[0115] Pooling Layer: Follows the convolutional layer immediately, used to reduce the spatial dimension of the feature map while retaining important information. Max Pooling is commonly used.

[0116] Fully - connected Layer: In the last few layers, flattens the feature map and further abstracts the feature representation through several fully - connected layers.

[0117] Dropout Layer: To prevent overfitting, a Dropout layer is added after some fully - connected layers to randomly discard some neurons.

[0118] Output Layer: Applies the Softmax function to convert it into a probability distribution, representing the prediction probability of each class.

[0119] A training module for training the neural network model;

[0120] When the training module trains the neural network model, the content used includes:

[0121] Supervised Learning Algorithm: Given a dataset with labels, let the neural network model learn the mapping relationship from input to output.

[0122] Loss function: The cross-entropy loss is used to measure the difference between the predicted distribution and the true distribution;

[0123] Optimizer: With good convergence speed and adaptive learning rate characteristics, it minimizes the loss function;

[0124] Regularization techniques: Including L2 regularization and early stopping method, which are used to avoid overfitting and improve the performance of the neural network model.

[0125] Test and validation modules for evaluating the neural network model.

[0126] The process of the test and validation modules for evaluating the neural network model is as follows:

[0127] Split the dataset: The original dataset is divided into a training set, a validation set, and a test set to ensure there are enough samples for evaluating the model's generalization ability;

[0128] Evaluation metrics: Four important metrics, accuracy, precision, recall, and F1-score, are used to measure the performance of the classification model, and the confusion matrix is used to analyze the misclassification situations between different classes, with attention paid to the detection accuracy of unqualified classes;

[0129] Continuous improvement: Adjust the hyperparameters according to the performance on the validation set, retrain the model, and finally evaluate the actual effectiveness of the model using the test set.

[0130] The welding quality detection process of the sheet metal parts of the present invention is as follows:

[0131] S1. Loading of sheet metal parts: The automated loading and unloading mechanism places the sheet metal parts to be detected on the conveyor belt;

[0132] S2. Image acquisition: The conveyor belt transports the sheet metal parts to the rotating detection table of the transparent glass, and a high-resolution camera takes surface images of the sheet metal parts from six angles;

[0133] S3. Real-time transmission: The images are immediately transmitted to the server for processing through a high-speed network interface;

[0134] S4. Data preprocessing: Under GPU acceleration, the images are quickly preprocessed, and the preprocessing includes operations such as size adjustment and normalization;

[0135] S5. Model inference: The pre-trained CNN model performs classification prediction on the processed images to obtain the probability values of each category;

[0136] S6. Decision inference: Determine whether the sheet metal parts are qualified according to the set threshold, and trigger an alarm mechanism for unqualified sheet metal parts;

[0137] S7. Feedback and recording: The system generates a report and stores the detection results for subsequent review and statistical analysis;

[0138] S8. Workpiece blanking: After inspection, the sheet metal parts are automatically moved out of the inspection area to continue the welding quality inspection of the next sheet metal part.

[0139] The core technology of the present invention is to utilize deep learning, especially the convolutional neural network (CNN) built based on the TensorFlow framework, to achieve the automatic inspection of the welding quality of sheet metal parts. The following is a detailed elaboration on this core technology:

[0140] Network architecture: Considering the characteristics of image data and the requirements of the inspection task, a convolutional neural network (CNN) suitable for processing two-dimensional image data is selected. CNN has characteristics such as local receptive fields, weight sharing, and pooling layers, which can effectively extract features in images and reduce the number of parameters, thereby improving the generalization ability and computational efficiency of the model.

[0141] Input and output definitions:

[0142] Input: Images of the welding parts of sheet metal parts taken from multiple angles by a high-resolution camera.

[0143] Output: The model will predict one of 10 categories, including "qualified" and 9 different types of "unqualified" (such as cracks, pores, incomplete penetration, etc.).

[0144] Preprocessing steps: Before the image is input into the network, a series of preprocessing operations are required, including resizing the image, image segmentation, normalizing pixel values, enhancing contrast, removing noise, etc., to ensure the consistency and high quality of the input data.

[0145] Training process: including

[0146] Dataset preparation: Collect and label a large number of images of sheet metal parts in different welding quality states as the training dataset. Each sample needs to be reviewed by three professional personnel to ensure the accuracy of the labels. To increase the robustness of the model, various factors that may affect the welding quality, such as different welding materials and process parameters, should also be included as much as possible.

[0147] Data augmentation: By applying random rotations, flips, scalings, etc. to the original images, more diverse training samples can be generated, which helps to improve the generalization ability of the model.

[0148] Model training: Use TensorFlow to build a CNN model and adopt the gradient descent method (such as the Adam optimizer) to minimize the loss function (such as cross-entropy loss). During the training process, the network weights are continuously updated through the backpropagation algorithm until the model converges or reaches the predetermined maximum number of iterations. In addition, an early stopping mechanism (EarlyStopping) can be introduced to prevent overfitting.

[0149] Verification and Tuning: Set aside a portion of the data as a validation set. Regularly evaluate the model's performance during training and adjust hyperparameters (such as learning rate, batch size, etc.) based on the validation results to find the optimal configuration.

[0150] Testing and Evaluation: Finally, evaluate metrics such as the accuracy, recall, and F1-score of the model on an independent test set to ensure it meets the requirements of practical applications.

[0151] Implementation of Classification and Alarm Functions

[0152] Classification Decision: The trained CNN model is deployed on the server side to receive real-time image streams from 6 high-resolution cameras. After each image is inferred by the model, a probability distribution vector is obtained, indicating the likelihood of belonging to each category. Set a threshold or select the category corresponding to the maximum probability as the final prediction result.

[0153] Alarm Trigger: Once the model determines that there is a certain form of defect in the workpiece (i.e., it does not belong to the "qualified" category), the system will immediately activate the alarm mechanism. This can be achieved in various ways, such as visual alarms (displaying red warning messages on the screen), sound alarms (playing a prompt tone through the speaker), or directly notifying the production line control system to stop operations for timely issue handling.

[0154] Feedback Loop: For products marked as unqualified, an additional manual review process can be set up to allow experienced engineers to further check whether the model's judgment is correct. If misclassifications are found, these new samples can be added to the training set to retrain the model to continuously improve the accuracy and reliability of the system.

[0155] In summary, through the combination of advanced automation control technology and deep learning algorithms, the present invention realizes efficient, accurate, and intelligent welding quality detection of sheet metal parts, significantly improving the quality management level of the manufacturing industry.

[0156] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A sheet metal welding quality detection device based on artificial intelligence and automatic control technology, characterized in that: include: Automatic loading and unloading mechanism: responsible for automatically transporting the sheet metal parts to be inspected to the inspection area, and moving the workpieces out of the inspection area after the inspection is completed; Conveying mechanism: used to transfer sheet metal parts from the automatic loading and unloading mechanism to the transparent glass-like rotating inspection table; High-resolution camera: used to shoot sheet metal parts on the rotating inspection table at multiple angles to obtain high-quality sheet metal part image data; Server: Receives and processes image data from high-resolution cameras and performs quality inspection using a pre-trained neural network model; The automatic loading and unloading mechanism and the rotating inspection platform are seamlessly connected at both ends of the conveying mechanism. The high-resolution camera is electrically connected to the server. A plurality of high-resolution cameras are arranged around the rotating inspection platform, and the high-resolution camera is configured to have high resolution, fast frame rate and good color reproduction.

2. The sheet metal welding quality detection device based on artificial intelligence and automatic control technology according to claim 1 is characterized in that: The automatic loading and unloading mechanism includes a grabbing robot arm driven by a high-precision servo motor. The load range of the grabbing robot arm is 50-200KG, the repeatability accuracy of the grabbing robot arm is ±0.05mm, and the grabbing robot arm is integrated with the control system through a communication interface.

3. The sheet metal welding quality detection device based on artificial intelligence and automatic control technology according to claim 1 is characterized in that: The conveying mechanism includes a belt conveyor, the width of the belt on the belt conveyor is 800mm-1500mm, and the feeding speed of the belt conveyor can be adjusted between 0.5m / s-3m / s by a frequency converter; There are six high-resolution cameras, including two left detection cameras, two right detection cameras, one top detection camera, and one back detection camera. The resolution of the high-resolution camera is ≥20MP, the frame rate of the high-resolution camera is ≥90fps, and the high-resolution camera is equipped with a fixed-focus or zoom lens, and the distortion of the lens is <1%.

4. The sheet metal welding quality detection device based on artificial intelligence and automatic control technology according to claim 1 is characterized in that: The server comprises: Processor: Use multi-core high-performance CPU, CPU main frequency ≥ 3.2GHz; Graphics card: includes 32GB HBM2 video memory to accelerate model reasoning; Memory: includes an SSD of no less than 1 TB for the operating system and temporary storage, and an NVMe SSD for array storage of training data sets.

5. A sheet metal welding quality detection system based on artificial intelligence and automated control technology, a sheet metal welding quality detection device based on artificial intelligence and automated control technology according to any one of claims 1 to 4, characterized in that: include: Hardware unit, including Automatic loading and unloading module that automatically transfers the sheet metal parts to be inspected from the production line to the inspection area and moves them out of the inspection area after the sheet metal parts are inspected; Transfer sheet metal parts from the automatic loading and unloading module to the conveying module on the rotary inspection table; An image acquisition module that takes multi-angle shots of sheet metal parts on a rotating inspection table; A server that receives and processes the image data acquired by the image acquisition module and detects the image data; and auxiliary modules, the auxiliary modules including lighting, protective covers, and sound and light alarms; The automatic loading and unloading module, the conveying mechanism, and the rotating detection platform are arranged in sequence, the image acquisition module is installed on the rotating detection platform, and the image acquisition module is electrically connected to the server; Software units, including A data preprocessing module for cleaning and labeling the acquired sheet metal image data; A neural network model for quality inspection of preprocessed sheet metal images; A training module for training neural network models; Testing and validation module for evaluating neural network models.

6. The sheet metal welding quality inspection system based on artificial intelligence and automatic control technology according to claim 5 is characterized in that: The workflow of the data preprocessing module includes: Data collection: Collect images of sheet metal parts from different welding processes, materials and environmental conditions; Data cleaning: including removing outliers, standardizing and normalizing, and resizing sheet metal images; Data augmentation: Apply random rotation, flipping, scaling, and cropping transformations to expand the data set and improve the generalization ability of the neural network model; Marking: The staff will check each sheet metal image, mark those with good welding quality as qualified, and those with poor welding quality as unqualified, and mark the reasons for the failure.

7. The sheet metal welding quality inspection system based on artificial intelligence and automatic control technology according to claim 5 is characterized in that: The neural network model includes: Input layer: responsible for receiving image data as the starting point of the neural network model. The input dimension depends on the size of the image. Hidden layers: including convolutional layer, pooling layer, fully connected layer, and Dropout layer; Output layer: Apply the Softmax function to convert it into a probability distribution, indicating the predicted probability of each category.

8. The sheet metal welding quality inspection system based on artificial intelligence and automatic control technology according to claim 5 is characterized in that: When the training module trains the neural network model, the content used includes: Supervised learning algorithm: given a labeled data set, let the neural network model learn the mapping relationship from input to output; Loss function: Use cross entropy loss to measure the difference between the predicted distribution and the true distribution; Optimizer: Uses good convergence speed and adaptive learning rate characteristics to minimize the loss function; Regularization techniques: including L2 regularization and early stopping, which are used to avoid overfitting and improve the performance of neural network models.

9. The sheet metal welding quality inspection system based on artificial intelligence and automatic control technology according to claim 5 is characterized in that: The process of evaluating the neural network model in the testing and verification module is as follows: Split the dataset: Divide the original dataset into training set, validation set, and test set to ensure that there are enough samples to evaluate the generalization ability of the model; Evaluation indicators: Use accuracy, precision, recall, and F1 score to measure the effect of the classification model, and use confusion matrix to analyze the misjudgment between categories, focusing on the detection accuracy of unqualified categories; Continuous improvement: Adjust hyperparameters based on the performance on the validation set, retrain the model, and finally use the test set to evaluate the actual performance of the model.

Images