A Clean Barrel Production Quality Inspection System and Method Based on Machine Vision and AI

Abstract

This invention provides a laminar flow jar production quality inspection system and method based on machine vision and AI, belonging to the field of industrial product quality inspection technology. The system comprises five units: data acquisition, processing, AI quality inspection model construction and training, AI quality inspection, and result feedback. These units work collaboratively via an industrial Ethernet connection. The data acquisition unit acquires raw data such as multi-view images; the data processing unit preprocesses and integrates the data; the AI ​​quality inspection model construction and training unit builds and trains a three-level feature extraction network model; the AI ​​quality inspection unit performs quality inspection based on the model; and the result feedback unit pushes results and communicates with the production equipment. The method covers steps including data acquisition, processing, model construction and training, quality inspection, and result feedback. This invention achieves efficient and accurate detection and feedback control of laminar flow jar production quality.

Description

Technical Field

[0001] This invention relates to the field of industrial product quality inspection technology, specifically to a method based on machine vision and artificial intelligence (AI) technology. It is suitable for real-time monitoring and control of product quality in cleanroom container manufacturing enterprises. Background Technology

[0002] A laminar flow container is an ultra-clean, high-purity reagent packaging container specifically designed for the semiconductor industry. It is primarily used for storing and transporting wet electronic chemicals, ensuring materials remain uncontaminated in an extremely pure environment. Made of high-density polyethylene (HDPE), it boasts excellent stability and corrosion resistance, effectively isolating external impurities.

[0003] In the cleanroom container manufacturing industry, traditional quality inspection methods have many drawbacks.

[0004] On the one hand, manual inspection is inefficient and its accuracy is difficult to guarantee. The production line for laminar flow containers operates at a high speed, making it difficult for manual visual inspection to conduct a comprehensive and detailed check of each product, easily overlooking minor defects such as tiny scratches and air bubbles. Furthermore, prolonged inspection work can easily lead to operator fatigue, thus affecting the accuracy of the test results.

[0005] On the other hand, traditional automated inspection methods lack the ability to analyze complex defects. Some simple automated inspection equipment can only detect specific types of defects. For the various complex defects that may occur during the production of cleanroom containers, such as crack propagation and accelerated deformation, which require analysis in conjunction with time-series data, traditional equipment often cannot effectively identify and judge them.

[0006] Furthermore, existing detection technologies struggle to provide real-time feedback and production adjustments. Even when product defects are detected, the information cannot be promptly relayed to the production management system and related equipment, hindering timely adjustments to the production process and resulting in more defective products and increased production costs.

[0007] Meanwhile, traditional testing methods have shortcomings in data processing and model optimization. They lack effective integration and analysis of large amounts of testing data, making it impossible to optimize testing models based on data, and thus difficult to adapt to constantly changing production processes and product quality requirements. Summary of the Invention

[0008] This invention aims to solve the problems of low efficiency and error-proneness in manual inspection, as well as to overcome the shortcomings of traditional automated inspection methods in analyzing complex defects. It also aims to change the situation where existing inspection technologies cannot provide real-time feedback and production adjustments, and to address the deficiencies of traditional inspection methods in data processing and model optimization.

[0009] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0010] A cleanroom drum production quality inspection system based on machine vision and AI is characterized by comprising a data acquisition unit, a data processing unit, an AI quality inspection model construction and training unit, an AI quality inspection unit, and a result feedback unit. The data acquisition unit is responsible for acquiring various raw data during the production process. The data processing unit preprocesses and integrates the raw data. The AI ​​quality inspection model construction and training unit is used to build and train a quality inspection model based on AI. The AI ​​quality inspection unit is used to perform quality judgment and defect identification based on the constructed quality inspection model. The result feedback unit pushes the quality inspection results to the production management system and related equipment in real time. All the above units are connected via an industrial Ethernet network. The output of the data acquisition unit is connected to the input of the data processing unit. The output of the data processing unit is connected to both the input of the AI ​​quality inspection model construction and training unit and the input of the AI ​​quality inspection unit. The output of the AI ​​quality inspection model construction and training unit is connected to the input of the AI ​​quality inspection unit. The output of the AI ​​quality inspection unit is connected to the input of the result feedback unit.

[0011] In a preferred embodiment of the present invention, the AI ​​quality inspection model construction and training unit includes a model architecture building device, a training parameter configuration device, a model training and optimization device, and a model evaluation device. The model architecture building device constructs a three-level feature extraction network comprising a bottom layer, a middle layer, and a high layer. The bottom layer network adopts a convolutional layer structure, containing three parallel convolutional branches with convolutional kernels of 3×3, 5×5, and 7×7, respectively. Each branch contains two convolutional layers with a stride of 1 and padding = same. The number of convolutional kernels is 64. Each convolutional layer is followed by a batch normalization layer and a LeakyReLU activation function. The branch output feature maps are concatenated. The middle layer network introduces an attention mechanism to enhance key defect features. The high layer network uses a bidirectional LSTM network to analyze the defect evolution trend and sets two parallel output branches for detection and classification. The training parameter configuration device sets the model training parameters. The model training and optimization device divides the training set and validation set for training and optimization. The model evaluation device uses an independent test set to evaluate the model performance and generate an evaluation report.

[0012] In a preferred embodiment of the present invention, the data acquisition unit includes a multi-view image acquisition device, a light source adjustment device, and a time-series data synchronization device; the multi-view image acquisition device consists of 6 sets of cameras, which are respectively installed at the inlet, forming station, trimming station, cleaning station, drying station, and outlet of the clean barrel production line; the light source adjustment device is correspondingly set to the multi-view image acquisition device, and each set includes 4 dimmable LED lights; the time-series data synchronization device synchronously controls the shooting timing of the multi-view image acquisition device through a trigger signal and is connected to the encoder of the production line.

[0013] In a preferred embodiment of the present invention, the data processing unit includes an image preprocessing device, a data annotation device, a data enhancement device, and a data integration and storage device; the image preprocessing device performs noise reduction, distortion correction, and image alignment on the original image data; the data annotation device adopts a semi-automatic annotation method; the data enhancement device performs geometric transformation, color transformation, and defect simulation on the annotated image data; and the data integration and storage device integrates the image data preprocessed by the image preprocessing device, the annotation information, and the corresponding time-series data into a structured dataset and stores it.

[0014] In a preferred embodiment of the present invention, the AI ​​quality inspection unit receives real-time preprocessed image data and time-series data output by the data processing unit via industrial Ethernet, performs format verification, removes invalid data, adds a unique identifier to valid data, inputs real-time image data into a three-level feature extraction network, outputs defect detection results and overall quality judgment, analyzes the defect changes of the same cleanroom container at different production stations, marks cleanroom containers with serious trend defects as "key concerns" and provides priority feedback.

[0015] A quality inspection method for cleanroom drum production based on machine vision and AI, characterized by the following steps:

[0016] Acquire various raw data during the production process;

[0017] Preprocessing and integrating the raw data;

[0018] A quality inspection model was built based on AI and trained on the model.

[0019] Quality judgment and defect identification are based on the established quality inspection model;

[0020] The quality inspection results are pushed to the production management system and related equipment in real time.

[0021] In a preferred embodiment of the present invention, the raw data includes external images of the clean barrels collected by a data acquisition device consisting of six sets of high-definition industrial cameras respectively installed at the inlet, forming station, trimming station, cleaning station, drying station and outlet of the clean barrel production line, as well as bottom images of the barrels collected by a bottom-view camera installed below the production line.

[0022] In a preferred embodiment of the present invention, the steps of constructing a quality inspection model based on AI and training the model include: constructing a three-level feature extraction network comprising a bottom layer, a middle layer, and a high layer; the bottom layer network uses three parallel convolutional branches with different kernel sizes and performs relevant processing before channel concatenation; the middle layer network introduces a channel attention mechanism to enhance features; and the high layer network uses a bidirectional LSTM network with two parallel output branches; setting model training parameters, dividing the training set and validation set for training and optimization, using an independent test set to evaluate model performance, and adjusting the network structure based on the evaluation results;

[0023] In a preferred embodiment of the present invention, the data processing is as follows: the original image data is sequentially subjected to noise reduction, distortion correction, and image alignment to form a 360° panoramic unfolded image; defect samples are labeled using a semi-automatic labeling method; the labeled image data is expanded and mixed with the original dataset at a ratio of 3:1; the preprocessed image data, labeling information, and corresponding time series data are integrated into a structured dataset and stored in the form of a table or database.

[0024] In a preferred embodiment of the present invention, the AI ​​quality inspection step includes: receiving real-time preprocessed image data and time-series data, performing format verification, removing invalid data, adding a unique identifier to valid data, inputting real-time image data into a three-level feature extraction network, outputting defect detection results and overall quality judgment, analyzing the defect changes of the same cleanroom container at different production stations, marking cleanroom containers with serious trend defects as "key concerns" and providing priority feedback.

[0025] The present invention has the following beneficial technical effects:

[0026] Through a unique three-level feature extraction network, various defects can be accurately identified. For example, the bottom layer network uses convolutional kernels of different sizes to capture defect features from minute to large scale, the middle layer network strengthens key defect features through attention mechanism, and the top layer network combines time series data to analyze the defect evolution trend, so as to achieve efficient and accurate detection of the quality of the clean barrel. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the logical structure of the ultra-clean barrel production quality inspection system based on machine vision and AI provided by the present invention.

[0028] Figure 2 This is a schematic diagram of the data processing of the underlying network of the present invention.

[0029] Figure 3 This is a schematic diagram of data processing in the middle layer network of this invention.

[0030] Figure 4 This is a schematic diagram of data processing in the high-level network of the present invention.

[0031] Figure 5 The flowchart shows the ultra-clean drum production quality inspection method based on machine vision and AI provided by the present invention. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0033] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0034] like Figure 1 As shown, this invention provides a cleanroom drum production quality inspection system based on machine vision and AI. The system's overall architecture consists of five units: a data acquisition unit, a data processing unit, an AI quality inspection model construction and training unit, an AI quality inspection unit, and a result feedback unit. The data acquisition unit is responsible for acquiring various raw data during the production process; the data processing unit preprocesses and integrates the raw data; the AI ​​quality inspection model construction and training unit builds and trains a quality inspection model based on AI; the AI ​​quality inspection unit performs quality judgment and defect identification based on the constructed quality inspection model; and the result feedback unit pushes the quality inspection results to the production management system and related equipment in real time, providing a basis for production adjustments.

[0035] All the above units are connected via an industrial Ethernet network. The output of the data acquisition unit is connected to the input of the data processing unit. The output of the data processing unit is connected to the input of the AI ​​quality inspection model construction and training unit and the input of the AI ​​quality inspection unit. The output of the AI ​​quality inspection model construction and training unit is connected to the input of the AI ​​quality inspection unit. The output of the AI ​​quality inspection unit is connected to the input of the result feedback unit.

[0036] In a preferred embodiment of the present invention, the data acquisition unit is used to acquire various types of raw data in the production process, and it includes the following sub-units: a multi-view image acquisition device, a light source adjustment device, and a time-series data synchronization device.

[0037] The multi-view image acquisition device consists of six sets of high-definition industrial cameras, installed at the inlet, forming station, trimming station, cleaning station, drying station, and outlet of the laminar flow tank production line. Each set of cameras includes three cameras with different focal lengths of 2mm, 5mm, and 10mm, arranged in a 120° ring around the outside of the laminar flow tank. More preferably, a bottom-view camera is installed below the laminar flow tank production line to capture images of the bottom of the tank.

[0038] The light source adjustment device is correspondingly set up with the multi-view image acquisition device. Each group includes four dimmable LED lights, two of which are 5500K color temperature ring shadowless lights installed inside the camera's ring array to provide uniform basic illumination; the other two are 6000K color temperature strip lights, which can be adjusted to produce oblique light to enhance the shadow contrast of defects such as scratches and dents. The brightness of the light source can be adjusted via PWM signals, automatically switching preset parameters according to the material of the cleanroom bucket (such as PP, PE) and its surface finish.

[0039] The time-series data synchronization device synchronously controls the shooting timing of the multi-view image acquisition device through trigger signals, ensuring that the timestamp deviation of the image data of the cleanroom bucket at different workstations and different viewpoints at the same time is within a certain threshold range, such as 5 milliseconds. Simultaneously, the time-series data synchronization device is connected to the encoder of the production line, associating the image data with the production cycle time of the cleanroom bucket (such as forming time and transmission speed) to generate a dataset containing time-series information.

[0040] The multi-view image acquisition device and the light source adjustment device are connected via signal lines. The light source adjustment device automatically adjusts the lighting parameters based on the image brightness value fed back by the camera. The timing data synchronization device is connected to the multi-view image acquisition device and the production line encoder via Ethernet, respectively, for timing synchronization control of data acquisition and production rhythm.

[0041] The data processing unit preprocesses and integrates the raw data, and includes: an image preprocessing device, a data annotation device, a data enhancement device, and a data integration and storage device.

[0042] The image preprocessing device receives the raw image data output by the multi-view image acquisition device, first performs noise reduction processing, then performs distortion correction to eliminate lens distortion and viewpoint deviation; finally, it performs image alignment, using the edge of the clean bucket opening and the axis of the bucket body as reference features, and maps images from different viewpoints to a unified coordinate system to form a 360° panoramic unfolded image of the clean bucket.

[0043] The data annotation device employs a semi-automatic annotation method. 10% of the defect samples (such as scratches, bubbles, dents, etc.) are manually annotated, and annotation files containing defect type, location coordinates, and dimensions are generated using tools such as LabelMe. For the remaining samples, automatic annotation is performed using a pre-trained basic detection model, followed by manual review and correction.

[0044] The data augmentation device expands the labeled image data by performing geometric transformations, color transformations, and defect simulations. The augmented dataset is mixed with the original dataset at a 3:1 ratio for subsequent model training.

[0045] The data integration and storage device integrates the image data preprocessed by the image preprocessing device, the annotation information, and the corresponding time-series data from the time-series data synchronization device into a structured dataset. Each data entry contains fields such as image ID, multi-view image path, defect annotation information, acquisition time, and production station, and is stored in the form of a table or database.

[0046] The input end of the image preprocessing device is connected to the output end of the multi-view image acquisition device, and its output end is connected to the data annotation device and the data enhancement device respectively; the output end of the data annotation device is connected to the input end of the data enhancement device; the output ends of the data enhancement device and the time-series data synchronization device are both connected to the data integration and storage device.

[0047] The AI ​​quality inspection model construction and training unit is used to build a quality inspection model based on AI and train the model. It includes the following sub-units: model architecture building device, training parameter configuration device, model training and optimization device, and model evaluation device.

[0048] The model architecture building device is used to construct a three-level feature extraction network consisting of a bottom layer, a middle layer, and a high layer.

[0049] Preferably, the underlying network adopts a convolutional layer structure, which includes three parallel convolutional branches, each branch having two convolutional layers and 64 convolutional kernels.

[0050] More preferably, the three parallel convolutional branches use convolutional kernels of three different sizes: 3×3, 5×5, and 7×7. The smaller 3×3 convolutional kernel is used to capture more subtle, localized features in the image, and can keenly perceive and extract relevant features for minute defects with obvious details, such as minor scratches that may exist on the surface of the cleanroom container. The 5×5 convolutional kernel is used to comprehensively consider information within a slightly larger local area, and can more comprehensively extract feature information for defects such as texture features or slightly larger microbubbles. The 7×7 convolutional kernel has a larger receptive field, and is used to integrate image information over a larger area, and has better feature extraction capabilities for features that appear in a larger area, such as large-area texture anomalies.

[0051] Within each branch, the number of convolutional layers is set to 2, with a stride of 1 for each layer and padding set to "same". A stride of 1 means that the convolutional kernel moves only one pixel at a time as it slides across the image, ensuring more precise feature extraction and preventing the loss of any possible details. Setting padding to "same" ensures that the size of the output feature map remains consistent with the size of the input image after the convolution operation.

[0052] Each branch has 64 convolution kernels, meaning that each branch extracts image features from 64 angles simultaneously during the convolution operation.

[0053] After each convolutional layer, a batch normalization (BN) layer is applied. The BN layer normalizes the mean and variance of the feature maps, adjusting them to a relatively stable distribution, making the network more stable during training. Following the BN layer, the LeakyReLU activation function is applied, with its slope preferably set to 0.1.

[0054] After the above operations, each of the three branches outputs a feature map with different feature emphases. To fully utilize the rich feature information extracted from these multi-scale processes, the outputs of these three branches need to be fused. Here, a channel concatenation operation is used, that is, the feature maps output by the three branches are concatenated along the channel dimension. Since each branch has 64 convolutional kernels, the concatenated feature map forms a feature map with 192 (64×3) channels. The fused feature map integrates the feature information extracted by the three convolutional kernels at different scales, including various basic features such as image edges and textures from fine to large scales, enabling the system to more accurately capture low-level defect features such as fine scratches and microbubbles on the surface of the cleanroom container.

[0055] Preferably, the middle-layer network, serving as an intermediate step in the three-level feature extraction network, enhances key defect features by introducing an attention mechanism based on the 192-channel feature map output by the bottom-layer network. See [link to diagram] for a schematic representation of the above process. Figure 2 .

[0056] The middle layer network includes a channel attention sub-device that assigns different weights to 192 feature channels, highlighting channel features related to defects and weakening irrelevant background channels. For example, it can highlight depth variation channels in recessed areas or color difference channels in color difference areas, while weakening uniform texture channels on the surface of the cleanroom container.

[0057] More preferably, the channel attention sub-device first performs global average pooling and global max pooling operations simultaneously on the input 192-channel feature map of size H×W×192.

[0058] The global average pooling method is to calculate the average value of all pixels in each channel, compress the H×W×1 channel features into a 1×1×1 scalar, and finally obtain a 1×1×192 feature vector, which reflects the overall strength of the channel features.

[0059] The global max pooling method calculates the maximum value of pixels in each channel, resulting in a 1×1×192 feature vector. This vector is more sensitive to local salient features within the channel, such as gradient abrupt changes at concave edges.

[0060] After global average pooling and global max pooling operations, the channel attention sub-device performs feature transformation and fusion. The channel attention sub-device inputs two 1×1×192 feature vectors into a shared fully connected layer network. This network contains one hidden layer (96 neurons), employs the ReLU activation function to introduce a non-linear transformation, and then maps the features back to the original channel dimension through the output layer (192 neurons). The two vectors are then processed by the fully connected layer and summed element-wise, fusing the global information from average pooling and the local information from max pooling.

[0061] More preferably, the channel attention sub-device further performs the following operations: It generates 1×1×192 channel attention weights from the fused feature vectors using an activation function. The generated channel weights are then element-wise multiplied with the original 192-channel feature map to achieve selective enhancement of different channels. The above processing is as follows: Figure 3 As shown.

[0062] High-level network: Combining the recurrent neural network (RNN) structure with a bidirectional LSTM network.

[0063] The high-level network, serving as the top layer of the three-level feature extraction network, is primarily used to analyze defect evolution trends. In a preferred embodiment of this invention, the high-level network employs a bidirectional LSTM network. First, the feature map output from the middle-layer network is flattened into sequence features along the spatial dimension, and then input into a two-layer bidirectional LSTM network. Each LSTM layer has 256 LSTM units, and Dropout is used to prevent overfitting. The output of the LSTM network is mapped to a temporal feature vector through a fully connected layer, used to analyze multi-frame image sequences and temporal data, extract defect evolution trend features, and identify high-level complex defects such as crack propagation and deformation aggravation.

[0064] Specifically, the higher-level network first performs feature preprocessing on the feature map output by the middle-level network. This is because the feature map output by the middle-level network has a size of H×W×192, which is a three-dimensional feature containing both spatial and channel information. To accommodate the sequential input requirements of the bidirectional LSTM network, this feature map needs to be preprocessed to convert it into sequential features.

[0065] The specific operation involves flattening the feature map along its spatial dimensions. This means expanding the H×W spatial dimension into a sequence of length H×W, where each sequence element corresponds to a 192-dimensional channel feature at a spatial location in the original feature map. In this way, the original three-dimensional feature map is transformed into a sequence feature of length H×W and dimension 192.

[0066] Then, the high-level network captures the temporal correlations of the sequence features. For defect detection in cleanroom containers, for example, when analyzing crack propagation, it is necessary not only to know how the crack extends from its initiation point in a certain direction, but also to understand the impact of feature changes in the surrounding area on its propagation. The high-level network processes the sequence features both forward and backward. The two-layer bidirectional LSTM network uses a layer-by-layer processing approach. The first layer performs preliminary processing on the input sequence features, extracting some basic temporal correlation features, and then passes these features to the second layer. The second layer performs deeper analysis and learning based on the processing results of the first layer, further mining more complex and abstract temporal correlation information in the sequence features.

[0067] For high-level networks, due to the need to handle complex defect evolution characteristics and the large number of network parameters, overfitting is prone to occur. Setting the Dropout probability to 0.3 means that approximately 30% of LSTM units will be randomly dropped during each training iteration, thus preventing the network from becoming overly reliant on the outputs of certain neurons. Those skilled in the art will understand that Dropout is a commonly used regularization technique in neural network training, primarily used to prevent model overfitting, and will not be elaborated upon further here.

[0068] Finally, the high-level network integrates and maps the features output by the bidirectional LSTM network, transforming them into a 128-dimensional temporal feature vector. This temporal feature vector contains the evolution characteristics of defects in the clean barrel in time and space, such as the speed and direction of crack propagation and the degree of deformation intensification.

[0069] More preferably, after the high-level network output, two parallel output branches are set up. The detection branch adopts the detection head structure of YOLOv5 and outputs the defect bounding box, confidence level and defect type, such as the classification probability of defects such as scratches, bubbles, dents, color difference, cracks, deformation, impurities, and interfaces. The classification branch outputs the overall qualified / unqualified judgment result of the clean barrel.

[0070] The training parameter configuration device is used to set relevant parameters for model training, including but not limited to learning rate, batch size, training epochs, optimizer, loss function, etc.

[0071] The model training and optimization device reads the dataset from the data integration and storage device and divides it into a training set and a validation set at a certain ratio, such as 8:2. Common transfer learning strategies and teacher models in the field can be used to train and optimize the model.

[0072] After training is completed, the model evaluation device uses an independent test set accounting for 10% of the total data to evaluate the model performance and generate an evaluation report to provide a basis for model optimization.

[0073] The outputs of the model architecture building device and the training parameter configuration device are both connected to the model training and optimization device. The output of the model training and optimization device is connected to the model evaluation device. The feedback results from the model evaluation device, such as identifying weak links in defect identification, can be returned to the model architecture building device to adjust the network structure, such as increasing the number of convolutional kernels corresponding to specific defects.

[0074] The AI ​​quality inspection unit is used to perform quality judgment and defect identification based on the constructed quality inspection model. In a preferred embodiment of the present invention, the AI ​​quality inspection unit receives real-time preprocessed image data output by the data processing unit and time-series data from the time-series data synchronization device via industrial Ethernet, performs format verification on the input data, removes invalid data, and adds a unique identifier (UUID) to the valid data.

[0075] The AI ​​quality inspection unit inputs real-time image data into a three-level feature extraction network. The bottom layer network outputs basic feature maps, the middle layer network strengthens key features through an attention mechanism, and the top layer network combines time-series data, such as a sequence of the last 5 frames of images, to perform sequence feature analysis. Finally, the classification and detection head outputs defect detection results and overall quality judgment. Based on the defect type classification probability, the specific type of each defect is determined, such as "minor scratches" or "circular bubbles," and the actual size and location of the defect are calculated by combining image coordinates, such as "the upper 1 / 3 area on the left side of the barrel."

[0076] The AI ​​quality inspection unit also analyzes the defect changes of the same laminar flow container at different production stations based on the time-series feature vector output by the high-level network. For example, it calculates the growth rate of crack length from forming to finishing stations, or analyzes the change curve of container deformation with transmission time. Laminar flow containers with serious trend defects are marked as "key concerns" and given priority feedback.

[0077] The result feedback unit pushes the quality inspection results to the production management system and related equipment in real time. In one illustrative embodiment, the result feedback unit integrates the defect information output by the AI ​​quality inspection unit, such as type, location, size, overall quality judgment (qualified / unqualified), and defect trend analysis results, such as "crack propagation risk," into a standardized JSON format report. This report includes fields such as a unique data identifier, the production number of the cleanroom container, the quality inspection time, and processing suggestions (e.g., "pressure adjustment at the repair station"). For unqualified products, the defect severity level is marked (minor / moderate / severe). The severity level is determined comprehensively based on the defect size, quantity, and trend. For example, a single defect with a diameter ≥2mm or a crack length ≥5mm is considered severe.

[0078] The result feedback unit pushes the quality inspection results to relevant nodes in different ways, such as to industrial displays, management personnel terminals, and audible and visual alarm devices.

[0079] More preferably, the result feedback unit establishes communication with the PLC system of the laminar flow jar production equipment, converting the processing suggestions in the quality inspection results into equipment control parameters. For three or more consecutive non-conforming products, a stop signal is sent to the production line, and the line is restarted after manual troubleshooting.

[0080] like Figure 5 As shown, another embodiment of the present invention provides a method for quality inspection of cleanroom drum production based on machine vision and AI, including the following steps:

[0081] Step 1: Data Acquisition. This involves acquiring various raw data from the production process. Specifically, this includes installing a data acquisition device consisting of six sets of high-definition industrial cameras at the entrance, forming station, trimming station, cleaning station, drying station, and exit of the laminar flow tank production line. Each set contains three cameras with focal lengths of 2mm, 5mm, and 10mm, arranged in a 120° ring around the outside of the laminar flow tank to capture external images. Simultaneously, a bottom-view camera is installed below the production line to capture images of the tank's bottom. Trigger signals are used to synchronously control the shooting timing of the multi-view image acquisition devices, ensuring that the timestamp deviation of the image data from different stations and perspectives at the same time is within 5 milliseconds. The data is also connected to the encoder of the production line to correlate the image data with the production cycle of the laminar flow tank, such as forming time and transmission speed, generating a dataset containing time-series information.

[0082] Step 2: Data Processing. Receive the raw image data acquired from multiple viewpoints, and sequentially perform noise reduction and distortion correction, including eliminating lens distortion and viewpoint deviation, and image alignment. Using the edge of the laminar flow hopper and the axis of the hopper body as reference features, images from different viewpoints are mapped to a unified coordinate system to form a 360° panoramic unfolded image.

[0083] A semi-automated annotation method is adopted. 10% of the samples with defects such as scratches, bubbles, and dents are manually annotated and annotated files containing defect type, location coordinates, and size are generated by annotation tools such as LabelMe. For the remaining samples, a pre-trained basic detection model is used for automatic annotation, and then manual review and correction are performed.

[0084] The labeled image data is augmented, including geometric transformations, color transformations, and defect simulations. The augmented dataset is then mixed with the original dataset at a 3:1 ratio for subsequent model training.

[0085] The preprocessed image data, annotation information, and corresponding time-series data are integrated into a structured dataset. Each data entry contains fields such as image ID, multi-view image path, defect annotation information, acquisition time, and production station, and is stored in the form of a table or database.

[0086] Step 3: AI Quality Inspection Model Construction and Training. This involves building and training an AI-based quality inspection model, specifically including:

[0087] Construct a three-level feature extraction network consisting of bottom, middle, and top layers:

[0088] The underlying network employs three parallel convolutional branches with kernel sizes of 3×3, 5×5, and 7×7, each containing two convolutional layers with a stride of 1 and padding of the same value. The number of kernels in each branch is 64. Each convolutional layer is followed by a batch normalization (BN) layer and a LeakyReLU activation function (slope 0.1). The feature maps output from the three branches are concatenated along the channel dimension to form a 192-channel feature map.

[0089] The middle layer network introduces a channel attention mechanism to perform global average pooling and global max pooling on the 192-channel feature map (both yielding 1×1×192 feature vectors). The two vectors are input into a shared fully connected layer network (containing one 96-neuron hidden layer and one 192-neuron output layer). After element-wise addition, the channel attention weights are generated through an activation function and multiplied with the elements of the original feature map to achieve feature enhancement.

[0090] High-level network: A bidirectional LSTM network is adopted. The output feature map of the middle layer is first flattened into sequence features (length H×W, dimension 192) according to the spatial dimension, and then input into a two-layer bidirectional LSTM network (256 units per layer, dropout probability 0.3). The output is mapped into a 128-dimensional temporal feature vector through a fully connected layer. Two parallel output branches are set. The detection branch uses the YOLOv5 detector head to output defect bounding boxes, confidence scores and types. The classification branch outputs the pass / fail judgment results.

[0091] Configure model training parameters, including learning rate, batch size, training epochs, optimizer, loss function, etc.

[0092] The integrated dataset was divided into training and validation sets in an 8:2 ratio, and training and optimization were performed using transfer learning strategies and a teacher model.

[0093] The model performance was evaluated using an independent test set representing 10% of the total data, and an evaluation report was generated. The network structure was then adjusted based on the evaluation results.

[0094] Step Four: AI Quality Inspection. Based on the constructed quality inspection model, quality judgment and defect identification are performed, specifically including:

[0095] Real-time preprocessed image data and time-series data are received via industrial Ethernet, format verification is performed, invalid data is removed, and a unique identifier (UUID) is added to valid data.

[0096] Real-time image data is input into a three-level feature extraction network. The bottom layer outputs basic feature maps, the middle layer strengthens key features through an attention mechanism, and the top layer performs sequence feature analysis by combining time-series data such as the sequence of the last 5 frames of images. The classification and detection head outputs defect detection results and overall quality judgment, and calculates the actual size and location of defects by combining image coordinates.

[0097] Based on the time-series feature vector of the high-level output, we analyze the defect changes of the same laminar flow container at different production stations, and mark laminar flow containers with serious trend defects as "key focus" and give them priority feedback.

[0098] Step 5: Results Feedback. Integrate defect information (type, location, size, overall quality assessment) and trend analysis results (such as "crack propagation risk") into a standardized JSON format report, including fields such as unique data identifier, cleanroom container production number, quality inspection time, and handling recommendations; for non-conforming products, determine the severity level based on defect size, quantity, and trend.

[0099] The quality inspection results are pushed to relevant nodes such as industrial displays, management terminals, and audible and visual alarm devices.

[0100] Establish communication with the PLC system of the laminar flow container production equipment and convert processing suggestions into equipment control parameters; for more than three consecutive non-conforming products, send a stop signal to the production line and restart after manual troubleshooting.

[0101] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0102] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A machine vision and AI-based ultra-clean bucket production quality inspection system, characterized in that, The system comprises a data acquisition unit, a data processing unit, an AI quality inspection model construction and training unit, an AI quality inspection unit, and a result feedback unit. The data acquisition unit is responsible for acquiring various types of original data in the production process. The data processing unit pre-processes and integrates the original data. The AI quality inspection model construction and training unit is used to construct a quality inspection model based on AI and train the model. The AI quality inspection unit is used to make quality judgments and defect identifications based on the constructed quality inspection model. The result feedback unit pushes the quality inspection results to the production management system and related equipment in real time. The above units are connected through an industrial Ethernet. The output end of the data acquisition unit is connected to the input end of the data processing unit. The output end of the data processing unit is connected to the input end of the AI quality inspection model construction and training unit and the input end of the AI quality inspection unit. The output end of the AI quality inspection model construction and training unit is connected to the input end of the AI quality inspection unit. The output end of the AI quality inspection unit is connected to the input end of the result feedback unit. The data acquisition unit comprises a multi-view image acquisition device, a light source adjustment device, and a time sequence data synchronization device. The multi-view image acquisition device comprises six sets of high-definition industrial cameras, which are installed at the entrance, forming station, trimming station, cleaning station, drying station, and exit of the super-clean barrel production line. Each set of cameras includes three cameras with different focal lengths, i.e., 2mm, 5mm, and 10mm. The three cameras with different focal lengths are arranged in a 120° ring shape outside the super-clean barrel. An overhead camera is arranged below the super-clean barrel production line. The AI quality inspection model construction and training unit comprises a model architecture building device, a training parameter configuration device, a model training and optimization device, and a model evaluation device. The model architecture building device constructs a three-level feature extraction network including a bottom layer, a middle layer, and a high layer. The bottom layer network adopts a convolutional layer structure, which includes three parallel convolution branches with convolution kernels of 3×3, 5×5, and 7×7, respectively. Each branch includes two convolution layers, each followed by a batch normalization layer and a LeakyReLU activation function. The branch output feature maps are concatenated. The middle layer network introduces an attention mechanism to strengthen key defect features. The high layer network adopts a bidirectional LSTM network to analyze the defect evolution trend and sets two parallel output branches for detection and classification. The training parameter configuration device sets the model training parameters. The model training and optimization device divides the training set and validation set for training and optimization. The model evaluation device evaluates the model performance using an independent test set and generates an evaluation report.

2. The machine vision and AI based ultra-clean bucket production quality inspection system according to claim 1, wherein, The light source adjustment device is correspondingly arranged with the multi-view image acquisition device, and each set includes four adjustable light-emitting diode (LED) lamps. The time sequence data synchronization device synchronously controls the shooting time of the multi-view image acquisition device through a trigger signal and is connected with an encoder of the production line. 3.The machine vision and AI based ultra-clean bucket production quality inspection system according to claim 2, wherein, The data processing unit comprises an image preprocessing device, a data labeling device, a data enhancement device, and a data integration storage device; the image preprocessing device performs noise reduction processing, distortion correction, and image alignment on original image data; the data labeling device adopts a semi-automatic labeling method; the data enhancement device performs geometric transformation, color transformation, and defect simulation on the labeled image data; and the data integration storage device integrates the preprocessed image data, labeling information, and corresponding time sequence data into a structured data set and stores it.

4. The machine vision and AI-based ultra-clean bucket production quality inspection system according to claim 3, wherein, The AI quality inspection unit receives the real-time preprocessed image data and time sequence data output by the data processing unit through an industrial Ethernet, performs format verification, eliminates invalid data, adds a unique identifier to the valid data, inputs the real-time image data into a three-level feature extraction network, outputs defect detection results and overall quality judgments, analyzes defect changes of the same super-clean barrel at different production stations, marks the super-clean barrel with serious trend defects as "focus" and prioritizes feedback.

5. A machine vision and AI-based ultra-clean bucket production quality inspection method, characterized in that, The method comprises the following steps: Various types of original data in the production process are acquired, including external images of the super-clean barrel collected by 6 sets of high-definition industrial cameras respectively arranged at the entrance, forming station, trimming station, cleaning station, drying station, and exit of the super-clean barrel production line, and a bottom image of the barrel collected by an overhead camera arranged below the production line, each set of the 6 sets of high-definition industrial cameras comprising 3 cameras with focal lengths of 2 mm, 5 mm, and 10 mm respectively, the 3 cameras being distributed in a 120° ring shape outside the super-clean barrel, the shooting time of the multi-view image collection device being controlled synchronously through a trigger signal to ensure that the image data time stamp deviation of the super-clean barrel at different stations and different angles at the same time is within 5 milliseconds, and the image data is associated with the production pace of the super-clean barrel production line to generate a data set containing time sequence information; The original data is preprocessed and integrated, the data processing comprising sequentially performing noise reduction processing, distortion correction, and image alignment on the original image data, taking the barrel mouth edge and barrel axis of the super-clean barrel as the reference features, mapping the images of different angles to a unified coordinate system to form a 360° panoramic development image, adopting a semi-automatic labeling method, manually labeling 10% of the defect samples, generating a labeling file containing defect types, position coordinates, and sizes through a labeling tool, automatically labeling the remaining samples by using a pre-trained basic detection model, and then manually auditing and correcting, expanding the labeled image data, including geometric transformation, color transformation, and defect simulation, mixing the expanded data set and the original data set at a ratio of 3:1, integrating the preprocessed image data, labeling information, and corresponding time sequence data into a structured data set, each piece of data containing image ID, multi-view image path, defect labeling information, acquisition time, production station, and other fields, and storing in the form of a table or a database; The AI quality inspection unit receives the real-time preprocessed image data and time sequence data output by the data processing unit through an industrial Ethernet, performs format verification, eliminates invalid data, adds a unique identifier to the valid data, inputs the real-time image data into a three-level feature extraction network, outputs defect detection results and overall quality judgments, analyzes defect changes of the same super-clean barrel at different production stations, marks the super-clean barrel with serious trend defects as "focus" and prioritizes feedback. The AI-based quality inspection model is constructed and trained, and a three-level feature extraction network including a bottom layer, a middle layer and a high layer is constructed. The bottom layer network adopts three parallel convolution branches with convolution kernel sizes of 3*3, 5*5 and 7*7 respectively, each branch contains 2 convolution layers with a step of 1 and padding=same, and the number of convolution kernels is 64. Each convolution layer is followed by a batch normalization layer and a LeakyReLU activation function with a slope of 0.

1. The feature maps output by the three branches are spliced in the channel dimension to form a 192-channel feature map. The middle layer network introduces a channel attention mechanism, and the 192-channel feature map is subjected to global average pooling and global maximum pooling, both of which obtain a 1*1*192 feature vector. The two vectors are input into a shared fully connected layer network containing a 96-neuron hidden layer and a 192-neuron output layer. After element addition, the channel attention weight is generated through an activation function, and the original feature map element is multiplied to realize feature enhancement. The high layer network adopts a bidirectional LSTM network. The output feature map of the middle layer is flattened into a sequence feature in the spatial dimension and input into a double-layer bidirectional LSTM network with 256 units per layer and a Dropout probability of 0.

3. The output is mapped to a 128-dimensional time sequence feature vector through a fully connected layer. Two parallel output branches are set up. The detection branch uses a YOLOv5 detection head to output defect bounding boxes, confidence and type. The classification branch outputs qualified / unqualified judgment results. Model training parameters are set, including learning rate, batch size, training rounds, optimizer, loss function, etc. The integrated stored dataset is divided into training set and validation set in the ratio of 8:

2. The model is trained and optimized using the transfer learning strategy and the teacher model. An independent test set of 10% of the total data is used to evaluate the model performance, and an evaluation report is generated. The network structure is adjusted according to the evaluation results. Based on the constructed quality inspection model, quality judgment and defect recognition are performed. Real-time preprocessed image data and time sequence data are received through industrial Ethernet, format verification is performed, invalid data is removed, and a unique identifier is added to the valid data. The real-time image data is input into the three-level feature extraction network. The bottom layer outputs a basic feature map, the middle layer strengthens key features through an attention mechanism, and the high layer analyzes sequence features combined with time sequence data such as the last 5 image sequences. The actual size and position of the defect are calculated based on the image coordinates. For the time sequence feature vector output by the high layer, the defect changes of the same super-clean bucket at different production stations are analyzed, and the super-clean bucket with serious trend defects is marked for special attention and priority feedback. The quality inspection result is pushed to the production management system and related equipment in real time, and the defect information and trend analysis result is integrated into a standardized JSON format report, including a data unique identifier, a super clean bucket production number, a quality inspection time, a processing suggestion and the like. For unqualified products, the severity level is determined according to the defect size, quantity and trend, and the quality inspection result is pushed to related nodes such as an industrial display, a management personnel terminal, an audible and visual alarm device and the like. Communication is established with a PLC system of the super clean bucket production equipment, the processing suggestion is converted into a device control parameter, and a stop signal is sent to the assembly line for more than three unqualified products in succession. After manual troubleshooting, the assembly line is restarted.

Images