Customized intelligent processing system based on MES system

Abstract

The invention relates to the technical field of intelligent quality inspection, and particularly discloses a customized intelligent processing system based on an MES (Manufacturing Execution System), which is characterized in that when a scanning device is used for reading two-dimensional code information on a plate, a camera is used for collecting a target plate image; a quality inspection module based on a deep learning algorithm is further introduced into the main control computer to process the image of the target plate, comprehensive perception and defect recognition of the surface state of the target plate are achieved through multi-level feature extraction and semantic alignment interactive fusion operation between features, and the defect recognition efficiency is improved for the plate with defects. And the quality inspection result is sent to the MES system for recording and subsequent processing, so that the MES system fully considers the quality condition of the plates during sorting decision making, and it is ensured that only qualified plates enter the next procedure. According to the invention, real-time monitoring and management of the plate quality can be realized, the production efficiency and the product quality are improved, and resource waste and rework cost caused by defective plates are reduced.

Description

Technical Field

[0001] This application relates to the technical field of intelligent quality inspection, and more specifically, to a customized intelligent processing system based on the MES system. Background Art

[0002] In modern manufacturing, especially in the field of customized furniture production, improving production efficiency and product quality is the key to the sustainable development of enterprises. In the traditional production process of customized panel furniture, the sorting and quality inspection of panels often rely on manual operations, which are not only inefficient but also prone to errors due to human factors, affecting the final quality of products. With the rapid development of information technology, especially the widespread application of MES (Manufacturing Execution System), automation and intelligence have become an important direction for the transformation and upgrading of the manufacturing industry.

[0003] For example, the invention patent with the publication number CN111097704B discloses an intelligent sorting system for customized panel furniture based on MES. It reads the two-dimensional code information on the panel through a code scanning device and sends it to the main control computer. After the main control computer obtains the panel information, it feeds it back to the MES system. The MES system gives the distribution instructions for the sorting channel and the shelf according to the panel information and the current sorting status. After receiving the distribution instructions from the MES system, the main control computer controls the corresponding sorting channel conveyor line to transport the panel to the corresponding sorting buffer area, thus realizing the automatic sorting of customized furniture panels.

[0004] However, the above solution mainly relies on two-dimensional codes for panel information reading and sorting path planning, and lacks effective automatic detection means for the actual quality status of the panel surface, such as defects like scratches, color differences, and deformations. As a result, defective panels may enter the subsequent processes, affecting the quality of the final product.

[0005] Therefore, an optimized customized intelligent processing system based on the MES system is needed. Summary of the Invention

[0006] To solve the above technical problems, the present application is proposed. An embodiment of the present application provides a customized intelligent processing system based on the MES system. While using a scanning device to read the two-dimensional code information on the board, it uses a camera to collect images of the target board, and further introduces a quality inspection module based on deep learning algorithms in the main control computer to process the images of the target board. Through multi-level feature extraction and semantic alignment and interaction fusion operations between features, it realizes a comprehensive perception of the surface state of the target board and defect recognition. For the board with defects, its quality inspection results are sent to the MES system for recording and subsequent processing, so that the MES system fully considers the quality status of the board when making sorting decisions, ensuring that only qualified boards enter the next process, and it can realize real-time monitoring and management of the board quality, improve production efficiency and product quality, and reduce resource waste and rework costs caused by defective boards.

[0007] Correspondingly, according to one aspect of the present application, a customized intelligent processing system based on the MES system is provided, which includes:

[0008] A loading elevator, a belt roller conveyor line, a code scanning device, a main control computer, and multiple sorting buffer areas. Each sorting buffer area is correspondingly provided with a sorting channel conveyor line and a shelf. The loading elevator is arranged at the entrance of the belt roller conveyor line; the code scanning device is arranged at the entrance of the belt roller conveyor line and is used to identify the board information of the board by scanning the code; the main control computer, the code scanning device, and the sorting channel conveyor line are communicatively connected, and are used to obtain the board information and feedback the board information to the MES system, and after obtaining the sorting channel and shelf allocation instructions fed back by the MES system, control the corresponding sorting channel conveyor line to transport the board on the corresponding sorting buffer area to the sorting discharge port, and generate board handling information to be carried. Among them, the main control computer further includes a quality inspection module for board quality inspection;

[0009] While obtaining the board information, the code scanning device uses a camera to collect images of the target board and sends the images of the target board to the quality inspection module. The quality inspection module is used to perform board defect detection based on the images of the target board to generate quality inspection results;

[0010] Among them, the board quality inspection module includes:

[0011] An image multi-scale feature extraction unit for performing multi-scale feature extraction on the images of the target board to obtain shallow features of the surface state of the target board and deep features of the surface state of the target board;

[0012] A feature alignment interaction unit for performing semantic alignment interaction between multi-scale features of the shallow features of the surface state of the target board and the deep features of the surface state of the target board to obtain a deep-shallow fine-grained alignment fusion feature of the surface state of the target board;

[0013] A quality inspection result generation unit for determining a quality inspection result based on the deep-shallow fine-grained alignment fusion feature of the surface state of the target board, where the quality inspection result is used to indicate whether there are defects on the surface of the target board;

[0014] A quality inspection result feedback unit for, in response to the quality inspection result indicating that there are defects on the surface of the target board, feeding back the quality inspection result to the MES system.

[0015] Preferably, the image multi-scale feature extraction unit is used for:

[0016] Inputting the target board image into a board multi-scale feature extractor based on the FPT model to obtain a shallow feature encoding vector of the surface state of the target board and a deep feature encoding vector of the surface state of the target board, respectively, as the shallow feature of the surface state of the target board and the deep feature of the surface state of the target board.

[0017] Preferably, the feature alignment interaction unit includes:

[0018] A semantic shift analysis subunit for performing semantic shift analysis on the shallow feature encoding vector of the surface state of the target board and the deep feature encoding vector of the surface state of the target board to obtain a deep-shallow fine-grained semantic information field of the surface state of the target board;

[0019] A semantic alignment encoding subunit for performing feature mapping semantic alignment interaction on the shallow feature encoding vector of the surface state of the target board and the deep feature encoding vector of the surface state of the target board based on the deep-shallow fine-grained semantic information field of the surface state of the target board to obtain a deep-shallow semantic alignment encoding vector of the surface state of the target board as the deep-shallow fine-grained alignment fusion feature of the surface state of the target board.

[0020] Preferably, the semantic shift analysis subunit includes:

[0021] A dimension modulation secondary subunit for inputting the shallow feature encoding vector of the surface state of the target board and the deep feature encoding vector of the surface state of the target board into a dimension modulation module based on point convolution to obtain a modulated shallow feature encoding vector of the surface state of the target board and a modulated deep feature encoding vector of the surface state of the target board, where the modulated shallow feature encoding vector of the surface state of the target board and the modulated deep feature encoding vector of the surface state of the target board have the same feature dimension;

[0022] The semantic information field encoding secondary subunit is used to perform fine-grained association encoding on the encoded vector of the shallow features of the surface state of the modulated target plate and the encoded vector of the deep features of the surface state of the modulated target plate, and then input it into the semantic information field encoding network based on the convolutional kernel to obtain the deep-shallow fine-grained semantic information field of the surface state of the target plate.

[0023] Preferably, the semantic information field encoding secondary subunit is used for:

[0024] Calculate the product of the encoded vector of the shallow features of the surface state of the modulated target plate and the transposed vector of the encoded vector of the deep features of the surface state of the modulated target plate, and then divide it by the square root of the feature scale value of the encoded vector of the deep features of the surface state of the modulated target plate to obtain the deep-shallow fine-grained association encoding matrix of the surface state of the target plate;

[0025] Use the semantic information field encoding network to perform multi-layer convolutional processing on the deep-shallow fine-grained association encoding matrix of the surface state of the target plate based on a 3×3 convolutional kernel to obtain the deep-shallow fine-grained semantic information field of the surface state of the target plate.

[0026] Preferably, the semantic alignment encoding subunit is used for:

[0027] Map the encoded vector of the shallow features of the surface state of the modulated target plate and the encoded vector of the deep features of the surface state of the modulated target plate to the deep-shallow fine-grained semantic information field of the surface state of the target plate respectively to obtain the fine-grained aligned encoded vector of the shallow features of the surface state of the target plate and the fine-grained aligned encoded vector of the deep features of the surface state of the target plate;

[0028] Calculate the position-wise weighted sum between the fine-grained aligned encoded vector of the shallow features of the surface state of the target plate and the fine-grained aligned encoded vector of the deep features of the surface state of the target plate to obtain the deep-shallow semantic alignment encoding vector of the surface state of the target plate.

[0029] Preferably, the quality inspection result generation unit is used for:

[0030] Input the deep-shallow fine-grained alignment fusion encoded vector of the surface state of the target plate into the defect detector based on the classifier to obtain the quality inspection result.

[0031] Preferably, the quality inspection result generation unit is used for:

[0032] Use the fully connected layer of the defect detector to perform fully connected encoding on the deep-shallow fine-grained alignment fusion encoded vector of the surface state of the target plate to obtain the deep-shallow fine-grained alignment fusion fully connected encoded vector of the surface state of the target plate;

[0033] Input the deep-shallow fine-grained alignment fusion full-connection encoding vector of the surface state of the target plate into the Softmax classification function of the defect detector to obtain the probability values of the deep-shallow fine-grained alignment fusion full-connection encoding vector of the surface state of the target plate belonging to each classification label, where the classification labels include qualified quality inspection and unqualified quality inspection;

[0034] Determine the classification label corresponding to the largest of the probability values as the quality inspection result.

[0035] The present application has at least the following technical effects:

[0036] Compared with the prior art, the customized intelligent processing system based on the MES system provided by the present application, while using the scanning device to read the two-dimensional code information on the plate, uses the camera to collect the image of the target plate, and further introduces a quality inspection module based on the deep learning algorithm in the main control computer to process the image of the target plate. Through multi-level feature extraction and semantic alignment and interaction fusion operations between features, it realizes the comprehensive perception and defect recognition of the surface state of the target plate, and for the plate with defects, sends its quality inspection result to the MES system for recording and subsequent processing, enabling the MES system to fully consider the quality status of the plate when making sorting decisions, ensuring that only qualified plates enter the next process. In this way, real-time monitoring and management of the plate quality can be achieved, production efficiency and product quality can be improved, and resource waste and rework costs caused by defective plates can be reduced. Description of the Drawings

[0037] By describing the embodiments of the present application in more detail in conjunction with the drawings, the above and other objects, features, and advantages of the present application will become more obvious. The drawings are used to provide a further understanding of the embodiments of the present application, and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation to the present application. In the drawings, the same reference numerals generally represent the same components or steps.

[0038] Figure 1 It is a block diagram of the plate quality inspection module in the customized intelligent processing system based on the MES system according to the embodiment of the present application.

[0039] Figure 2 It is a schematic diagram of the data flow of the plate quality inspection module in the customized intelligent processing system based on the MES system according to the embodiment of the present application.

[0040] Figure 3 It is a block diagram of the feature alignment and interaction unit in the customized intelligent processing system based on the MES system according to the embodiment of the present application.

[0041] Figure 4It is a block diagram of a semantic offset analysis sub-unit in a customized intelligent processing system based on an MES system according to an embodiment of the present application. Detailed implementation manners

[0042] As shown in the present application and the claims, unless the context clearly indicates otherwise, words such as "a", "an", "one", and / or "the" are not specifically singular and may also include the plural. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of the steps and elements that have been clearly identified, and these steps and elements do not constitute an exclusive list. The method or device may also include other steps or elements.

[0043] Although the present application makes various references to certain modules in the system according to the embodiments of the present application, however, any number of different modules can be used and run on the user terminal and / or the server. The modules are only illustrative, and different aspects of the system and method can use different modules.

[0044] In the present application, flowcharts are used to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the operations before or below are not necessarily performed precisely in sequence. On the contrary, as needed, various steps can be performed in reverse order or simultaneously. At the same time, other operations can also be added to these processes, or one or several operations can be removed from these processes.

[0045] Next, exemplary embodiments of the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. It should be understood that the present application is not limited by the exemplary embodiments described herein.

[0046] It is worth noting that in the present application, all actions of obtaining data are carried out on the premise of complying with the corresponding data protection regulations and policies of the country where it is located and obtaining the authorization given by the owner of the corresponding device.

[0047] As described in the above background art, Patent CN111097704B proposes an intelligent sorting system for panel customized furniture based on MES, which includes: a loading elevator, a belt roller conveyor line, a scanning device, a main control computer, and multiple sorting buffer areas. Each of the sorting buffer areas is correspondingly provided with a sorting channel conveyor line and a shelf. The loading elevator is arranged at the entrance of the belt roller conveyor line; the scanning device is arranged at the entrance of the belt roller conveyor line for identifying the panel information of the panel by scanning; the main control computer, the scanning device, and the sorting channel conveyor line are communicatively connected, used for obtaining the panel information and feeding back the panel information to the MES system, and after obtaining the sorting channel and shelf allocation instructions fed back by the MES system, controlling the corresponding sorting channel conveyor line to transport the panel arriving at the corresponding sorting buffer area to the sorting discharge port, and generating the panel handling pending information.

[0048] The above intelligent sorting system for panel customized furniture based on MES realizes the automatic sorting of panels through the coordinated work of components such as the loading elevator, the belt roller conveyor line, the scanning device, the main control computer, and the sorting buffer areas, improving the sorting efficiency and accuracy. However, since the above solution mainly relies on two-dimensional codes for panel information reading and sorting path planning, and lacks effective automatic detection means for the actual quality status of the panel surface (such as defects like scratches, color differences, deformations, etc.), defective panels may enter the subsequent processes, thereby affecting the quality of the final product.

[0049] To address this problem, based on the above solution, this application further introduces a quality inspection module for panel quality inspection into the main control computer. At the same time, during the process of using the scanning device to obtain the panel information, a camera is also used to collect the target panel image to capture the detailed visual information on the surface of the target panel, and the collected target panel image is sent to the quality inspection module for panel defect detection to generate the corresponding quality inspection results. Among them, based on the deep learning algorithm, the quality inspection module can achieve a comprehensive perception of the surface state of the target panel and defect recognition through multi-level feature extraction and semantic alignment and interactive fusion operations between features on the target panel image. For the panel with defects, its quality inspection results are sent to the MES system for recording and subsequent processing, enabling the MES system to fully consider the quality status of the panel when making sorting decisions, ensuring that only qualified panels enter the next process, realizing real-time monitoring and management of the panel quality, thereby further improving production efficiency and product quality, and reducing resource waste and rework costs caused by defective panels.

[0050] In the production process of customized panels, the scanning device is set at the entrance of the belt roller conveyor line. This location selection ensures that every panel entering the production line can have its information read and identified in a timely manner. The main function of the scanning device is to identify the panel information of the panel by scanning, and this process provides key data support for a series of subsequent operations.

[0051] The technical principle adopted by the scanning device is based on optical character recognition (OCR) or two-dimensional code / barcode reading technology. When the panel reaches the entrance of the belt roller conveyor line, the two-dimensional code or barcode pre-printed or attached on the panel will enter the working range of the scanning device. The scanning device uses the built-in optical sensor to capture the information of these codes and converts the graphic information into a data format that can be read by a computer through a decoding algorithm. In this way, all relevant information on the panel, including but not limited to dimensions, material type, color, surface treatment requirements, and special processing instructions specified by the customer, can be accurately read and transmitted to the main control computer. To ensure the efficiency and accuracy of the scanning process, the scanning device is usually equipped with a high-resolution camera and advanced image processing software. Even on a high-speed production line, it can achieve stable and accurate information reading of the fast-moving panel. In addition, the scanning device also has a certain fault tolerance ability and can still maintain a high recognition success rate under adverse conditions such as light changes and deviation of the panel placement angle. For situations where information cannot be read normally, such as damage or stain coverage of the two-dimensional code, the scanning device can immediately send an alarm signal to the main control system. This prompts the system to take corresponding measures, such as pausing the current process to check the problem, or guiding the panel to the manual processing station for separate processing. This immediate feedback mechanism helps to reduce the possibility of errors flowing into the next process, thus maintaining the smooth operation of the entire production line.

[0052] Meanwhile, a stable communication connection has been established between the barcode scanning device and the main control computer. As the central brain of the system, the main control computer is responsible for coordinating the data exchange between the barcode scanning device and the sorting channel conveyor line. Through modern communication technologies such as industrial Ethernet or wireless networks, the main control computer can establish stable data links with the barcode scanning devices and sorting channel conveyor lines distributed at various positions on the production line in real time. This communication connection method ensures the speed and reliability of information transmission, enabling efficient and stable operation even in complex and changing manufacturing environments. Once the panel information is read, the barcode scanning device will immediately send the acquired data to the main control computer. After receiving the panel information transmitted by the barcode scanning device, the main control computer will first conduct a preliminary verification of these data to ensure that all necessary information is complete and correct. Then, according to the pre-set rules and logic, the main control computer will classify and sort this information, and calculate the optimal sorting path and storage plan in combination with factors such as the production plan and inventory status of the MES system. This process involves complex algorithm operations and big data analysis, aiming to maximize production efficiency and minimize resource waste.

[0053] After determining the specific sorting strategy, the MES system will generate detailed sorting channel and shelf allocation instructions and feedback them to the main control computer through a secure and reliable communication protocol (such as OPC UA, MQTT, etc.). To ensure the accuracy of instruction execution, a strict verification mechanism is built into the main control computer. This mechanism can not only verify whether the instruction format from the MES system is correct, but also check whether the instruction content is logical and whether there are potential conflicts. If any abnormal situation is found, the main control computer will immediately issue a warning to the MES system and request to resend the correct instruction. In addition, the main control computer will record the process of each instruction reception and parsing, forming a log file for subsequent auditing and problem tracing.

[0054] After completing the instruction parsing, the main control computer will start the control program for the sorting channel conveyor line. Each sorting channel conveyor line is equipped with an independent drive motor and control system, which can accept the operation commands issued by the main control computer. According to the allocation instructions of the MES system, the main control computer will precisely set the speed, direction, and other necessary operating parameters of each conveyor line to ensure that the panel can move smoothly along the predetermined path. To cope with possible emergencies such as panel jams or equipment failures, various types of protection devices and emergency response mechanisms are installed on the conveyor line, which can detect problems and take measures to solve them in the first time. For example, hardware facilities such as emergency stop buttons and overload protection switches, combined with intelligent diagnostic software, can quickly locate the fault point without affecting the overall production process and shorten the downtime.

[0055] The sorting buffer area serves as a temporary storage place for the panels and plays a buffering role during the entire sorting process, enabling panels of different batches to queue up orderly for further processing. Each sorting buffer area is connected to a specific sorting channel conveyor line and is equipped with automated elevators or other auxiliary equipment to adjust the height of the panels so that they can smoothly enter the buffer area. When a panel arrives at the sorting buffer area, the system will verify the panel's identity information again to ensure that it indeed belongs to the target object in this area. This is an important step to avoid confusion in subsequent processes caused by incorrect allocation. At the same time, various monitoring facilities are also installed inside the buffer area, such as RFID tag readers, video surveillance cameras, etc., which can track dynamic information such as the quantity and location of the panels in real time, providing the main control computer with the latest status reports to adjust the production rhythm in a timely manner.

[0056] Next, the main control computer needs to continue to command the relevant equipment to transport the panels from the sorting buffer area to the sorting discharge port. To achieve this, the system will activate the conveying device in the sorting buffer area at an appropriate time to push the panels one by one onto the transition conveyor line connected to the sorting discharge port. This conveyor line is also closely monitored by the main control computer and can flexibly adjust the speed and direction according to the actual situation to ensure that the panels pass through the discharge port safely and stably. In addition, to improve efficiency, intelligent scheduling algorithms can be adopted, which can plan the best path in advance according to the destinations of different panels to avoid unnecessary waiting time.

[0057] When the panel is about to leave the sorting system, the main control computer will automatically generate a detailed panel handling information. This information includes the basic attributes of the panel (such as size, material, color), processing requirements (such as cutting, drilling, surface treatment), and the next logistics arrangement (such as which workstation to be sent to, estimated arrival time).

[0058] Figure 1 It is a block diagram of the panel quality inspection module in the customized intelligent processing system based on the MES system according to the embodiment of the present application. Figure 2 It is a schematic diagram of the data flow of the panel quality inspection module in the customized intelligent processing system based on the MES system according to the embodiment of the present application. As Figure 1 and Figure 2As shown in the figure, the plate quality inspection module 100 includes: an image multi-scale feature extraction unit 110, configured to perform multi-scale feature extraction on the target plate image to obtain a shallow feature of the target plate surface state and a deep feature of the target plate surface state; a feature alignment interaction unit 120, configured to perform semantic alignment interaction between multi-scale features of the shallow feature of the target plate surface state and the deep feature of the target plate surface state to obtain a deep-shallow fine-grained alignment fusion feature of the target plate surface state; a quality inspection result generation unit 130, configured to determine a quality inspection result based on the deep-shallow fine-grained alignment fusion feature of the target plate surface state, where the quality inspection result is used to indicate whether there are defects on the target plate surface; and a quality inspection result feedback unit 140, configured to, in response to the quality inspection result indicating that there are defects on the target plate surface, feedback the quality inspection result to the MES system.

[0059] In the above customized intelligent processing system based on the MES system, the image multi-scale feature extraction unit 110 is configured to perform multi-scale feature extraction on the target plate image to obtain a shallow feature of the target plate surface state and a deep feature of the target plate surface state. In a specific example of the present application, the image multi-scale feature extraction unit 110 is configured to: input the target plate image into a plate multi-scale feature extractor based on the FPT model to obtain a shallow feature encoding vector of the target plate surface state and a deep feature encoding vector of the target plate surface state, respectively, as the shallow feature of the target plate surface state and the deep feature of the target plate surface state. Specifically, the present application takes into account various types of defects that may exist on the plate surface, such as local detail defects such as scratches and stains, as well as overall structure defects such as color difference and deformation. Therefore, in order to achieve comprehensive detection of different types of defects, the present application uses the FPT model (Feature Pyramid Transformer) to construct a plate multi-scale feature extractor to perform multi-level plate surface feature extraction on the target plate image to obtain a shallow feature encoding feature vector of the target plate surface state and a deep feature encoding vector of the target plate surface state. Those of ordinary skill in the art should know that the FPT model is a deep learning model that combines a Feature Pyramid Network (FPN) and a Transformer architecture. Among them, the FPN network can effectively extract different scale feature information of the target plate image through a combination of bottom-up and top-down paths; while the Transformer architecture can capture the interdependence between features globally by applying a self-attention mechanism to each scale feature, enhancing the information expression ability of each scale feature. In this way, local detail information and global structure information on the target plate surface can be effectively captured, and comprehensive detection of various defects on the plate surface can be achieved.

[0060] In the above-mentioned customized intelligent processing system based on the MES system, the feature alignment interaction unit 120 is used to perform semantic alignment interaction between multi-scale features of the shallow features of the surface state of the target plate and the deep features of the surface state of the target plate to obtain the deep-shallow fine-grained alignment fusion features of the surface state of the target plate. It should be understood that in order to comprehensively utilize the multi-level feature information of the target plate image to achieve a comprehensive perception of the surface state of the plate, the present application further performs feature fusion on the encoded feature vector of the shallow features of the surface state of the target plate and the encoded vector of the deep features of the surface state of the target plate to generate a more complete and accurate representation of the surface state of the plate. In particular, considering that in the above-mentioned feature extraction process, multi-level feature extraction will cause information imbalance between shallow features and deep features, resulting in an information gap when high-level semantic information and low-level semantic information are simply cascaded. In response to this, the present application proposes a method for semantic alignment interaction between multi-scale features, which constructs a semantic information field between the two by learning the feature differences between the shallow features of the surface state of the target plate and the deep features of the surface state of the target plate, so that the shallow features and the deep features can be aligned and interacted at the semantic level. Among them, Figure 3 is a block diagram of the feature alignment interaction unit in the customized intelligent processing system based on the MES system according to an embodiment of the present application. As Figure 3 shown, the feature alignment interaction unit 120 includes: a semantic offset analysis subunit 121, which is used to perform semantic offset analysis on the encoded vector of the shallow features of the surface state of the target plate and the encoded vector of the deep features of the surface state of the target plate to obtain the deep-shallow fine-grained semantic information field of the surface state of the target plate; a semantic alignment encoding subunit 122, which is used to perform feature mapping semantic alignment interaction on the encoded vector of the shallow features of the surface state of the target plate and the encoded vector of the deep features of the surface state of the target plate based on the deep-shallow fine-grained semantic information field of the surface state of the target plate to obtain the deep-shallow semantic alignment encoded vector as the deep-shallow fine-grained alignment fusion features of the surface state of the target plate.

[0061] Figure 4 is a block diagram of the semantic offset analysis subunit in the customized intelligent processing system based on the MES system according to an embodiment of the present application. As Figure 4As shown, the semantic offset analysis sub-unit 121 includes: a dimension modulation secondary sub-unit 1211, configured to input the shallow feature encoding vector of the surface state of the target plate member and the deep feature encoding vector of the surface state of the target plate member into a dimension modulation module based on point convolution to obtain a modulated shallow feature encoding vector of the surface state of the target plate member and a modulated deep feature encoding vector of the surface state of the target plate member, wherein the modulated shallow feature encoding vector of the surface state of the target plate member and the modulated deep feature encoding vector of the surface state of the target plate member have the same feature dimension; a semantic information field encoding secondary sub-unit 1212, configured to perform fine-grained association encoding on the modulated shallow feature encoding vector of the surface state of the target plate member and the modulated deep feature encoding vector of the surface state of the target plate member, and then input the encoded vector into a semantic information field encoding network based on a convolution kernel to obtain the deep-shallow fine-grained semantic information field of the surface state of the target plate member.

[0062] More specifically, the dimension modulation secondary sub-unit 1211 is expressed by the formula:

[0063] v′1 = Leaky ReLU{Conv 1×1 (v1)}

[0064] v′2 = Leaky ReLU{Conv 1×1 (v2)}

[0065] Wherein, v1 represents the shallow feature encoding vector of the surface state of the target plate member, v2 represents the deep feature encoding vector of the surface state of the target plate member, Conv 1×1 (·) represents a point convolution operation, Leaky ReLU is a leaky rectified linear unit function, v′1 represents the modulated shallow feature encoding vector of the surface state of the target plate member, and v′2 represents the modulated deep feature encoding vector of the surface state of the target plate member.

[0066] That is, dimension modulation is performed by performing point convolution processing on the shallow feature encoding vector of the surface state of the target plate member and the deep feature encoding vector of the surface state of the target plate member to ensure that they have the same feature dimension and achieve preliminary dimension alignment.

[0067] More specifically, in a specific example of the present application, the semantic information field encoding secondary sub-unit 1212 is configured to: first, calculate the product of the modulated shallow feature encoding vector of the surface state of the target plate member and the transposed vector of the modulated deep feature encoding vector of the surface state of the target plate member, and then divide it by the square root of the feature scale value of the modulated deep feature encoding vector of the surface state of the target plate member to obtain a deep-shallow fine-grained association encoding matrix of the surface state of the target plate member, which is expressed by the formula:

[0068]

[0069] in,(·) T represents the transpose of a vector, represents a matrix multiplication operation, L is the characteristic scale value of the shallow characteristic coding vector of the surface state of the target plate after modulation and the deep characteristic coding vector of the surface state of the target plate after modulation, M x Represents the deep-shallow fine-grained semantic association encoding matrix of the target panel surface state.

[0070] Then, the semantic information field encoding network is used to perform multi-layer convolution processing based on a 3×3 convolution kernel on the deep-shallow fine-grained association encoding matrix of the target panel surface state to obtain the deep-shallow fine-grained semantic information field of the target panel surface state, which is expressed as follows:

[0071] Ω=Conv 3×3 (M x ) Among them, Conv 3×3 (·) represents a 3×3 convolution operation, and Ω represents the deep-shallow fine-grained semantic information field of the target panel surface state.

[0072] That is, by performing fine-grained associative coding on the shallow feature coding vector of the modulated target panel surface state and the deep feature coding vector of the modulated target panel surface state, the subtle semantic relationship between the two is captured, and a semantic information field is constructed accordingly. Specifically, the semantic information field coding network stacks multiple convolutional layers to further extract the deep and shallow correlation features of the target panel, so as to mine the position alignment and semantic offset information between the shallow feature coding vector of the target panel surface state and the deep feature coding vector of the target panel surface state, and gradually learn the mapping from low-level features to high-level semantic concepts, thereby constructing a deep-shallow fine-grained semantic information field of the target panel surface state, providing precise guidance for subsequent feature mapping.

[0073] Specifically, the semantic alignment coding subunit 122 is used to: first, map the shallow feature coding vector of the modulated target panel surface state and the deep feature coding vector of the modulated target panel surface state to the deep-shallow fine-grained semantic information field of the target panel surface state respectively to obtain the shallow feature coding vector of the fine-grained aligned target panel surface state and the deep feature coding vector of the fine-grained aligned target panel surface state, which is expressed by the formula:

[0074]

[0075] Among them, v 1t and v 2t They respectively represent the shallow feature encoding vector of the surface state of the fine-grained alignment target panel and the deep feature encoding vector of the surface state of the fine-grained alignment target panel.

[0076] That is, cross-domain feature transformation is performed on the shallow feature encoding vector of the surface state of the target plate after the above-mentioned dimensional modulation and the deep feature encoding vector of the surface state of the target plate, so as to map the two to the deep-shallow fine-grained semantic information field of the surface state of the target plate respectively, and achieve semantic alignment between the two. During the mapping process, each position feature in the modulated shallow feature encoding vector of the surface state of the target plate and the modulated deep feature encoding vector of the surface state of the target plate will be adjusted according to the semantic intensity and relevance of the corresponding position in the deep-shallow fine-grained semantic information field of the surface state of the target plate, so as to better match the semantic distribution in the semantic information field, thereby obtaining a fine-grained aligned shallow feature encoding vector of the surface state of the target plate and a fine-grained aligned deep feature encoding vector of the surface state of the target plate. In this way, not only the original semantic information of the shallow feature of the surface state of the target plate and the deep feature of the surface state of the target plate is retained, but also the two have the characteristic of fine-grained alignment of semantic information, that is, the semantic consistency between them is enhanced.

[0077] Then, calculate the position-wise weighted sum between the fine-grained aligned shallow feature encoding vector of the surface state of the target plate and the fine-grained aligned deep feature encoding vector of the surface state of the target plate to obtain the deep-shallow semantic alignment encoding vector of the surface state of the target plate, which is expressed by the formula:

[0078] v c = αv 1t + βv 2t

[0079] where α and β are learnable weight parameters, and v c represents the deep-shallow semantic alignment encoding vector of the surface state of the target plate.

[0080] That is, through the calculation method of position-wise weighted sum, the semantic interaction and fusion between the shallow feature of the surface state of the target plate and the deep feature of the surface state of the target plate are realized, and the deep-shallow fine-grained aligned fusion feature of the surface state of the target plate is generated, so as to comprehensively integrate multi-level feature information and more comprehensively reflect the surface state of the plate, providing a more accurate feature basis for subsequent defect detection.

[0081] In the above-mentioned customized intelligent processing system based on the MES system, the quality inspection result generation unit 130 is used to determine the quality inspection result based on the deep-shallow fine-grained aligned fusion feature of the surface state of the target plate, and the quality inspection result is used to indicate whether there are defects on the surface of the target plate. In a specific example of the present application, the quality inspection result generation unit 130 is used to: input the deep-shallow fine-grained aligned fusion encoding vector of the surface state of the target plate into a defect detector based on a classifier to obtain the quality inspection result.

[0082] More specifically, in a specific example of the present application, the quality inspection result generation unit 130 is configured to: perform a fully connected encoding on the deep-shallow fine-grained alignment fusion encoding vector of the surface state of the target plate using the fully connected layer of the defect detector to obtain a deep-shallow fine-grained alignment fusion fully connected encoding vector of the surface state of the target plate; input the deep-shallow fine-grained alignment fusion fully connected encoding vector of the surface state of the target plate into the Softmax classification function of the defect detector to obtain the probability values of the deep-shallow fine-grained alignment fusion fully connected encoding vector of the surface state of the target plate belonging to each classification label, where the classification labels include qualified quality inspection and unqualified quality inspection; determine the classification label corresponding to the largest of the probability values as the quality inspection result. Specifically, the defect detector first performs a fully connected encoding on the deep-shallow fine-grained alignment fusion encoding vector of the surface state of the target plate through its internal fully connected layer, mapping the data to a new feature space, which can be more abstract and discriminative while retaining most of the information of the original features, preparing for the subsequent classification task. Then, through the Softmax classification function, each element is converted into a real number between 0 and 1, representing the possibility that the plate belongs to the corresponding classification label. Finally, the classification label corresponding to the element with the largest probability is selected as the quality inspection result for output.

[0083] Considering that the shallow feature encoding feature vector of the surface state of the target plate and the deep feature encoding vector of the surface state of the target plate respectively represent the image semantic encoding features of different depths and different scales of the target plate image, during the alignment interaction between features based on the semantic information field, the deep-shallow fine-grained alignment fusion encoding vector of the surface state of the target plate may have probability convergence and divergence due to differences in feature depth and feature scale, affecting the accuracy of the quality inspection result obtained by the defect detector based on the classifier.

[0084] Preferably, inputting the deep-shallow fine-grained alignment fusion encoding vector of the surface state of the target plate into the defect detector based on the classifier to obtain the quality inspection result includes:

[0085] Inputting the deep-shallow fine-grained alignment fusion encoding vector of the surface state of the target plate into the defect detector based on the classifier to obtain a deep-shallow fine-grained alignment fusion encoding probability characterization quantity, where the deep-shallow fine-grained alignment fusion encoding probability characterization quantity is the probability value that the quality inspection result is qualified;

[0086] Performing a probability mapping on the feature values at each position in the deep-shallow fine-grained alignment fusion encoding vector of the surface state of the target plate based on the Sigmoid activation function to obtain a deep-shallow fine-grained alignment fusion probability encoding vector;

[0087] Based on the deep-shallow fine-grained alignment fusion probability coding vector of the surface state of the target plate and the deep-shallow fine-grained alignment fusion coding probability characterization quantity of the surface state of the target plate, calculate the deep-shallow fine-grained alignment fusion coding gradient parameter of the surface state of the target plate, expressed as:

[0088]

[0089] where, v i represents the eigenvalue at the i-th position of the deep-shallow fine-grained alignment fusion probability coding vector of the surface state of the target plate, p represents the deep-shallow fine-grained alignment fusion coding probability characterization quantity of the surface state of the target plate, and α represents the deep-shallow fine-grained alignment fusion coding gradient parameter;

[0090] Perform statistical analysis on the deep-shallow fine-grained alignment fusion coding vector of the surface state of the target plate to obtain the feature mean and feature variance, and apply the deep-shallow fine-grained alignment fusion coding gradient parameter of the surface state of the target plate to the feature mean and the feature variance respectively to obtain the first deep-shallow fine-grained alignment fusion coding distribution field mapping parameter and the second deep-shallow fine-grained alignment fusion coding distribution field mapping parameter of the surface state of the target plate, expressed as:

[0091] δ1 = μα

[0092] δ2 = σ 2 α

[0093] where, μ and σ 2 respectively represent the feature mean and feature variance of the deep-shallow fine-grained alignment fusion coding vector of the surface state of the target plate, δ1 represents the first deep-shallow fine-grained alignment fusion coding distribution field mapping parameter of the surface state of the target plate, and δ2 represents the second deep-shallow fine-grained alignment fusion coding distribution field mapping parameter of the surface state of the target plate;

[0094] Use the first deep-shallow fine-grained alignment fusion coding distribution field mapping parameter and the first deep-shallow fine-grained alignment fusion coding distribution field mapping parameter of the surface state of the target plate as non-linear scaling factors respectively to perform response characterization calculation on the deep-shallow fine-grained alignment fusion coding vector of the surface state of the target plate to obtain the first deep-shallow fine-grained alignment fusion coding fine-grained response characterization vector and the second deep-shallow fine-grained alignment fusion coding fine-grained response characterization vector of the surface state of the target plate, expressed as:

[0095]

[0096] Where, ⊙ represents dot product by position, V represents the deep-shallow fine-grained alignment fusion coding vector of the surface state of the target plate, exp represents the natural exponential function value, V1 represents the fine-grained response characterization vector of the deep-shallow fine-grained alignment fusion coding of the surface state of the first target plate, and V2 represents the fine-grained response characterization vector of the deep-shallow fine-grained alignment fusion coding of the surface state of the second target plate.

[0097] Dynamically optimize and fuse the fine-grained response characterization vector of the deep-shallow fine-grained alignment fusion coding of the surface state of the first target plate and the fine-grained response characterization vector of the deep-shallow fine-grained alignment fusion coding of the surface state of the second target plate to obtain an optimized deep-shallow fine-grained alignment fusion coding vector of the surface state of the target plate, denoted as:

[0098]

[0099] Where, ε represents the first weight hyperparameter, θ represents the second weight hyperparameter, represents vector addition, and V′ represents the optimized deep-shallow fine-grained alignment fusion coding vector of the surface state of the target plate;

[0100] Input the optimized deep-shallow fine-grained alignment fusion coding vector of the surface state of the target plate into a defect detector based on a classifier to obtain an identification result.

[0101] Correspondingly, by calculating the gradient attribute of the target sequence probability distribution corresponding to the deep-shallow fine-grained alignment fusion coding vector of the surface state of the target plate and inputting this attribute into the decentralized local feature fusion, a joint mapping of the distribution domain features of this vector is constructed based on the heterogeneous core modeling architecture. Furthermore, by enhancing the adaptability of the refined pattern representation to the category probability modeling, the coupling correlation relationship of the multi-level features of this vector in the global semantic space is regulated to strengthen the non-linear conversion from the high-dimensional space to the probability domain in the feature-category probability mapping optimization process, so as to ensure that the refined pattern representation and the category probability distribution continuously maintain a steady state balance of collaborative iteration on the mapping link from the feature domain to the category probability. In this way, the convergence effectiveness of the feature set with probability convergence and divergence of the deep-shallow fine-grained alignment fusion coding vector of the surface state of the target plate to the probability density distribution space is realized, and the accuracy of the quality inspection result obtained by the deep-shallow fine-grained alignment fusion coding vector of the surface state of the target plate through a defect detector based on a classifier is improved.

[0102] In the above-mentioned customized intelligent processing system based on the MES system, the quality inspection result feedback unit 140 is used to feedback the quality inspection result to the MES system in response to the quality inspection result indicating that there are defects on the surface of the target board. In actual applications, the main control computer communicates with the MES system through an interface and sends the quality inspection result to the MES system in the form of a data packet. After receiving the quality inspection result, the MES system will adjust the production process according to preset logical rules. For example, if the quality inspection result shows that there are defects on the surface of the target board, the MES system may trigger an alarm to notify the operator to conduct an inspection, or automatically adjust the equipment parameters on the production line to reduce similar defects that may occur in the future.

[0103] For the convenience of management and auditing, all quality inspection data will be stored in a central database, which supports historical query and data analysis. To ensure the security and integrity of the data, encryption technology and backup mechanisms are also adopted in the quality inspection reports. Each time a new quality inspection record is generated, the system will encrypt it to prevent unauthorized access or tampering. At the same time, all quality inspection data is regularly backed up to an offline storage device to prevent accidental loss. This dual protection measure ensures that even in the event of a cyber attack or hardware failure, important information can be retained to the greatest extent possible.

[0104] To ensure that the quality inspection results can be timely and accurately feedback to the MES system, a stable and reliable communication mechanism is established. This communication mechanism is based on industrial standard protocols, such as OPC UA, MQTT or HTTP / REST API, to ensure the security and reliability of data transmission. According to the API documentation of the MES system, corresponding client or server programs are developed to send and receive quality inspection results. These programs have good fault tolerance and exception handling mechanisms and can operate normally under network fluctuations or other unexpected situations. Message queue technologies (such as RabbitMQ, Kafka) are introduced as an intermediate layer between the quality inspection results and the MES system to effectively handle the data flood in high-concurrency scenarios, ensuring that each quality inspection message can be properly processed and will not be lost or delayed due to excessive instantaneous traffic. A series of monitoring metrics are set to track the performance parameters during the feedback process of the quality inspection results in real time. Once abnormal situations, such as communication interruption, response timeout, etc., are detected, an alarm is immediately triggered to notify relevant personnel to intervene.

[0105] When the MES system receives the quality inspection results, it immediately initiates a series of predefined operational processes to handle this situation. Based on the current status of the production line, the MES system adjusts the production plan to prevent defective panels from being included in subsequent processes. Specifically, the MES system queries the current production orders and schedules to identify qualified panels that can be substituted, or adjusts the speed and sequence of the production line to minimize the impact on the overall production rhythm. At the same time, the MES system updates inventory management, marking these panels as "pending processing" status to facilitate appropriate repair or disposal measures. In addition, the MES system notifies relevant workstations or operation terminals, providing specific instructions on how to handle this batch of defective panels. For situations that require manual intervention, the MES system can also generate work orders and assign them to responsible personnel for recheck or repair. Throughout the process, the MES system continuously monitors the production progress to ensure that all adjustments can be implemented smoothly and minimize the impact on the overall production rhythm.

[0106] In addition, for possible complex situations, the system is designed with a multi-level redundancy mechanism and a fault recovery plan. For example, when the primary communication link fails, the backup link will automatically take over the data transmission task to ensure uninterrupted information transfer; for problems that cannot be resolved for a long time, the system will trigger a higher-level alarm to prompt management to intervene and take emergency measures. At the same time, the system is built with a self-diagnosis function that can automatically detect potential problems during daily operation and provide early warnings. In this way, the system can always maintain a high level of reliability and stability.

[0107] In summary, the customized intelligent processing system based on the MES system according to the embodiments of the present application is elucidated. While using a scanning device to read the QR code information on the panel, it uses a camera to collect images of the target panel, and further introduces a quality inspection module based on deep learning algorithms in the main control computer to process the images of the target panel. Through multi-level feature extraction and semantic alignment and interactive fusion operations between features, it realizes a comprehensive perception of the surface state of the target panel and defect identification. For panels with defects, their quality inspection results are sent to the MES system for recording and subsequent processing, enabling the MES system to fully consider the quality status of the panels when making sorting decisions, ensuring that only qualified panels enter the next process. In this way, real-time monitoring and management of the panel quality can be achieved, improving production efficiency and product quality, and reducing resource waste and rework costs caused by defective panels.

[0108] The basic principles of the present invention have been described in conjunction with specific embodiments. However, it should be noted that the advantages, benefits, effects, etc. mentioned in the present invention are merely examples and not limitations. It cannot be considered that these advantages, benefits, effects, etc. are essential for each embodiment of the present invention. Additionally, the specific details of the above embodiments are only for the purpose of illustration and facilitating understanding, rather than limitations. The above details do not limit the present invention to necessarily adopt the above specific details for implementation.

[0109] In the above embodiments, the descriptions of each embodiment have their own emphases. For parts not detailed or recorded in a certain embodiment, reference can be made to the relevant descriptions of other embodiments. In the several embodiments provided by the present invention, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative. For example, the unit division is only a logical function division, and there can be other division methods in actual implementation. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they can be located in one place or distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0110] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and the present invention can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, it is intended to cover all changes falling within the meaning and scope of the equivalent elements of the claims in the present invention. Any associated drawing marks in the claims should not be regarded as limiting the claimed rights.

Claims

1. A customized intelligent processing system based on the MES system, comprising: A loading elevator, a belt roller conveyor line, a barcode scanning device, a main control computer, and a plurality of sorting buffer areas. Each sorting buffer area is correspondingly provided with a sorting channel conveyor line and a shelf. The loading elevator is arranged at the entrance of the belt roller conveyor line; the barcode scanning device is arranged at the entrance of the belt roller conveyor line and is used to identify the board information of the board by scanning the barcode; the main control computer, the barcode scanning device, and the sorting channel conveyor line are communicatively connected, and are used to obtain the board information and feedback the board information to the MES system, and after obtaining the sorting channel and shelf allocation instructions fed back by the MES system, control the corresponding sorting channel conveyor line to transport the board on the corresponding sorting buffer area to the sorting discharge port, and generate board handling required information. It is characterized in that the main control computer further includes a quality inspection module for board quality inspection; While obtaining the board information, the barcode scanning device uses a camera to collect a target board image and sends the target board image to the quality inspection module, and the quality inspection module is used to perform board defect detection based on the target board image to generate a quality inspection result; Among them, the board quality inspection module includes: An image multi-scale feature extraction unit for performing multi-scale feature extraction on the target board image to obtain a target board surface state shallow feature and a target board surface state deep feature; A feature alignment and interaction unit for performing semantic alignment and interaction between multi-scale features of the target board surface state shallow feature and the target board surface state deep feature to obtain a target board surface state deep-shallow fine-grained alignment and fusion feature; A quality inspection result generation unit for determining a quality inspection result based on the target board surface state deep-shallow fine-grained alignment and fusion feature, and the quality inspection result is used to indicate whether there are defects on the target board surface; A quality inspection result feedback unit for, in response to the quality inspection result being that there are defects on the target board surface, feeding back the quality inspection result to the MES system.

2. The customized intelligent processing system based on the MES system according to claim 1, characterized in that, The image multi-scale feature extraction unit is used for: Inputting the target board image into a board multi-scale feature extractor based on the FPT model to obtain a target board surface state shallow feature coding vector and a target board surface state deep feature coding vector respectively as the target board surface state shallow feature and the target board surface state deep feature.

3. The customized intelligent processing system based on the MES system according to claim 2, characterized in that, The feature alignment and interaction unit includes: A semantic shift analysis sub-unit for performing semantic shift analysis on the target board surface state shallow feature coding vector and the target board surface state deep feature coding vector to obtain a target board surface state deep-shallow fine-grained semantic information field; A semantic alignment coding sub-unit for, based on the target board surface state deep-shallow fine-grained semantic information field, performing feature mapping semantic alignment and interaction on the target board surface state shallow feature coding vector and the target board surface state deep feature coding vector to obtain a target board surface state deep-shallow semantic alignment coding vector as the target board surface state deep-shallow fine-grained alignment and fusion feature.

4. The customized intelligent processing system based on the MES system according to claim 3, characterized in that, The semantic shift analysis sub-unit includes: The dimension modulation secondary subunit is used to input the shallow feature encoding vector of the surface state of the target plate and the deep feature encoding vector of the surface state of the target plate into a dimension modulation module based on point convolution to obtain a modulated shallow feature encoding vector of the surface state of the target plate and a modulated deep feature encoding vector of the surface state of the target plate, wherein the modulated shallow feature encoding vector of the surface state of the target plate and the modulated deep feature encoding vector of the surface state of the target plate have the same feature dimension; The semantic information field encoding secondary subunit is used to perform fine-grained association encoding on the modulated shallow feature encoding vector of the surface state of the target plate and the modulated deep feature encoding vector of the surface state of the target plate, and then input them into a semantic information field encoding network based on a convolution kernel to obtain the deep-shallow fine-grained semantic information field of the surface state of the target plate.

5. The customized intelligent processing system based on the MES system according to claim 4, characterized in that, The semantic information field encoding secondary subunit is used for: Calculating the product of the modulated shallow feature encoding vector of the surface state of the target plate and the transposed vector of the modulated deep feature encoding vector of the surface state of the target plate, and then dividing by the square root of the feature scale value of the modulated deep feature encoding vector of the surface state of the target plate to obtain a deep-shallow fine-grained association encoding matrix of the surface state of the target plate; Using the semantic information field encoding network to perform multi-layer convolution processing on the deep-shallow fine-grained association encoding matrix of the surface state of the target plate based on a 3×3 convolution kernel to obtain the deep-shallow fine-grained semantic information field of the surface state of the target plate.

6. The customized intelligent processing system based on the MES system according to claim 5, wherein The semantic alignment encoding subunit is used for: Mapping the modulated shallow feature encoding vector of the surface state of the target plate and the modulated deep feature encoding vector of the surface state of the target plate to the deep-shallow fine-grained semantic information field of the surface state of the target plate respectively to obtain a fine-grained aligned shallow feature encoding vector of the surface state of the target plate and a fine-grained aligned deep feature encoding vector of the surface state of the target plate; Calculating the position-wise weighted sum between the fine-grained aligned shallow feature encoding vector of the surface state of the target plate and the fine-grained aligned deep feature encoding vector of the surface state of the target plate to obtain the deep-shallow semantic alignment encoding vector of the surface state of the target plate.

7. The customized intelligent processing system based on the MES system according to claim 6, wherein, The quality inspection result generation unit is used for: Inputting the deep-shallow fine-grained alignment fusion encoding vector of the surface state of the target plate into a defect detector based on a classifier to obtain the quality inspection result.

8. The customized intelligent processing system based on the MES system according to claim 7, wherein, The quality inspection result generation unit is used for: Performing fully connected encoding on the deep-shallow fine-grained alignment fusion encoding vector of the surface state of the target plate using the fully connected layer of the defect detector to obtain a deep-shallow fine-grained alignment fusion fully connected encoding vector of the surface state of the target plate; Inputting the deep-shallow fine-grained alignment fusion fully connected encoding vector of the surface state of the target plate into the Softmax classification function of the defect detector to obtain the probability values that the deep-shallow fine-grained alignment fusion fully connected encoding vector of the surface state of the target plate belongs to each classification label, wherein the classification labels include qualified quality inspection and unqualified quality inspection; Determining the classification label corresponding to the largest probability value as the quality inspection result.

Images