LED production control method and system based on machine vision guidance

Abstract

This invention provides a machine vision-guided LED production control method and system. It acquires images of each stage of LED production, obtaining original LED production images and corresponding equipment operating data and environmental sensor data. The original LED production images undergo super-resolution reconstruction to obtain high-resolution LED production images. A dynamic benchmark feature library for LED production is constructed by combining preset LED production process standard data. Defect features are extracted and identified from the high-resolution LED production images according to an LED production defect detection mechanism, generating a defect feature set. Error source tracing and correction mechanisms are used to trace the source of errors in the defect feature set. An error correction strategy is generated by combining the corresponding equipment operating data and environmental sensor data. Production parameter adjustment instructions are generated based on the error correction strategy and sent to the corresponding LED production control equipment for closed-loop control, thereby improving LED production quality and efficiency.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and more specifically, to a machine vision-guided LED production control method and system. Background Technology

[0002] As a highly efficient and energy-saving solid-state light source, LEDs are widely used in lighting, display and other fields. With the market's increasing requirements for product precision and reliability, their production quality control faces severe challenges. Production involves multiple stages such as die bonding and wire bonding. Deviations in any stage of the process may lead to product defects, thereby affecting optical and electrical performance and lifespan.

[0003] Traditional methods rely on manual inspection and experience-based parameter adjustments. On the one hand, manual inspection is inefficient and highly subjective; on the other hand, equipment parameters are mostly adjusted based on historical experience and cannot be dynamically optimized. Furthermore, when changes in the production environment or equipment status lead to product defects, parameter calibration is time-consuming, resulting in poor real-time performance. In addition, LED production data is mostly scattered, which often limits the correlation analysis of LED production data, resulting in weak traceability and closed-loop control capabilities for LED product quality. Summary of the Invention

[0004] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a machine vision-guided LED production control method, the method comprising: Image acquisition is performed on each stage of LED production to obtain original LED production images and corresponding equipment operating condition data and environmental sensor data. Super-resolution reconstruction processing is performed on the original LED production images to obtain high-resolution LED production images, and a dynamic benchmark feature library for LED production is constructed by combining it with preset LED production process standard data. Based on the LED production defect detection mechanism, defect features are extracted and identified from high-resolution LED production images to generate a defect feature set. Based on the error source tracing and correction mechanism, the defect feature set is used to trace the source of error, and error correction strategy is generated by combining the corresponding equipment operating condition data and environmental sensing data. The error correction strategy generates production parameter adjustment instructions, which are then sent to the corresponding LED production control equipment for closed-loop control.

[0005] As a further aspect of the present invention, super-resolution reconstruction processing is performed on the original LED production image to obtain a high-resolution LED production image, including: A generative adversarial network with embedded spatial attention and channel attention mechanisms is used as the super-resolution reconstruction model, and the original LED production image is input into the super-resolution reconstruction model. After the original LED production image is input into the model, multi-scale feature extraction is performed through a seven-layer cascaded convolutional neural network. Simultaneously, feature enhancement and cross-layer fusion reconstruction are performed based on the residual dense network. Based on the adversarial training mechanism, and combined with preset structural parameters and loss function, iterative optimization is performed to generate high-resolution LED production images.

[0006] As a further aspect of the present invention, a dynamic benchmark feature library for LED production is constructed by combining preset LED production process standard data, including: Analyze the preset LED production process standard data, extract the key quality parameters of each production link in the LED production process, such as die bonding, wire bonding, packaging and aging test, and generate the corresponding quality indicators for each production link. Based on the quality indicators, feature extraction is performed on the high-resolution LED production images corresponding to each production stage in the qualified production batch to generate standard feature templates for each production stage. Based on the standard feature template, a corresponding defect feature sub-library is constructed for each production stage. Each defect feature sub-library contains a defect feature vector database, the process parameter range range and quality judgment threshold for the corresponding production stage, and assigns a time tag and batch information to each defect feature. Among them, time tags and batch information are used to dynamically update the dynamic benchmark feature library for LED production.

[0007] As a further aspect of the present invention, defect features are extracted and identified from high-resolution LED production images based on an LED production defect detection mechanism to generate a defect feature set, including: A defect detection model is constructed by training a generative adversarial network based on federated learning and meta-learning. High-resolution LED production images are input into the defect detection model to identify defects and generate corresponding defect probability heatmaps. A non-maximum suppression algorithm with an adaptive threshold is used to filter defect regions in the defect probability heatmap to obtain candidate defect regions; Defect features are extracted from candidate defect regions to generate defect feature vectors, and the defect feature vectors are compared with the corresponding defect feature vector database in the LED production dynamic benchmark feature library. Based on the comparison results, the defect type and severity level of the candidate defect area are labeled, and the spatial coordinate information of the defect in the high-resolution LED production image is recorded simultaneously. The defect type, severity level and spatial coordinate information are then encapsulated to generate a defect feature set.

[0008] As a further aspect of the present invention, an error source tracing and correction mechanism is used to trace the source of errors in the defect feature set, and an error correction strategy is generated by combining the corresponding equipment operating condition data and environmental sensor data, including: Based on the LED production dynamic benchmark feature library, the defect types corresponding to defects generated during the LED production process are obtained, and the equipment operating condition data and environmental sensor data with the highest correlation are preset for different defect types. With the defect type as the root node, the corresponding equipment operating data and environmental sensor data as intermediate nodes, and the error source as the leaf node, an error source tracing decision tree is constructed. The defect feature set and the corresponding equipment operating condition data and environmental sensing data are input into the error tracing decision tree. Based on the defect type contained in the defect feature set, the tree is matched and located to the corresponding branch node at the end of the error tracing decision tree. The equipment operating condition data and environmental sensing data corresponding to the defect type are adjusted, the changes in the defect features corresponding to the defect type are analyzed, and the error source corresponding to the defect feature set is located. An initial error correction decision is generated based on the error source, and an error correction strategy is generated by combining the corresponding equipment operating data and environmental sensor data.

[0009] As a further aspect of the present invention, an initial error correction decision is generated based on the error source, and an error correction strategy is generated by combining the corresponding equipment operating condition data and environmental sensor data, including: An error source tracing report is generated based on the error source. An LED production state space is constructed based on the defect feature set, the error source tracing report, equipment operating data, and environmental sensor data. Error correction is performed based on reinforcement learning. The error tracing report should include at least the type of error source, the degree of impact on product quality, the associated equipment operating data and environmental sensor data, and the corresponding deviation value. The reinforcement learning agent, based on a preset reward function, learns by continuously interacting with equipment operating data and environmental sensor data within the LED production state space, generating targeted error correction strategies.

[0010] As a further aspect of the present invention, a production parameter adjustment instruction is generated according to an error correction strategy and sent to the corresponding LED production control equipment for closed-loop control, including: The error correction strategy is converted into digital control instructions that can be directly executed by the corresponding LED production equipment, and the timestamps of the digital control instructions are synchronized using a time-sensitive network-based timestamp synchronization mechanism. Before the instruction is sent, the digital control instruction is validated for its rationality based on the instruction rationality verification mechanism; For digital control commands that have passed the rationality verification, the digital control command is directly sent to the corresponding LED production control equipment. For digital control instructions that fail the rationality verification, return to the error correction strategy generation stage for modification and adjustment, and re-verify the rationality of the digital control instructions corresponding to the modified error correction strategy. Real-time monitoring of changes in production images corresponding to the execution of digital control commands by LED production control equipment, evaluation of the effects of the changes in production images, and generation of a closed-loop control report for LED production that includes command content, execution process, and correction effects; When the effect of N consecutive corrections is satisfactory, save the current model parameters and production parameters as an optimization template.

[0011] As a further aspect of the present invention, when the effect of N consecutive corrections is satisfactory, the corresponding current tuning parameters and production parameters are saved as an optimization template, including: The N value is set differently based on the production batch and product type corresponding to LED production. When no serious quality defects are found in N consecutive test results, and the rate of minor quality defects is lower than the preset rate of minor defects, the test sequence is judged to be qualified. Among them, serious quality defects are functional defects that directly affect the product's function and service life, while minor quality defects are surface defects that do not affect the product's function and service life. An optimization template is generated based on the tuning parameters and production parameters corresponding to the qualified inspection sequence, and the optimization template is matched with the standard feature templates corresponding to each defect feature sub-library in the LED production dynamic benchmark feature library; If a standard feature template is successfully matched, the standard feature template is then optimized based on the optimized template. If no standard feature template is matched, a defect feature sub-library is constructed based on the optimized template; Among them, the tuning parameters refer to the parameters corresponding to the super-resolution reconstruction processing, defect feature extraction and identification, and error tracing operations, while the production parameters refer to the optimal combination of equipment operating data and environmental sensor data corresponding to the qualified inspection sequence.

[0012] As a further aspect of the present invention, image acquisition is performed on each stage of LED production to obtain original LED production images and corresponding equipment operating condition data and environmental sensor data, including: The LED production process is synchronously captured using a multispectral camera array to obtain raw LED production images. The acquisition frequency is dynamically adjusted according to the production speed. Real-time operating data of the production equipment is collected synchronously by the equipment operating condition sensor group. The equipment operating condition data includes at least the welding head temperature, welding pressure, die bonding accuracy, motor running speed and conveyor belt position. The environmental sensor group collects environmental sensing data corresponding to the LED production environment. The environmental sensing data includes at least the ambient temperature, ambient humidity, ambient air pressure, air dust concentration, and light intensity. Data preprocessing is performed on raw LED production images, equipment operating data, and environmental sensor data. The data preprocessing includes at least spatiotemporal alignment and normalization operations.

[0013] Furthermore, embodiments of the present invention also provide an LED production control system based on machine vision guidance, comprising: The data acquisition module is used to acquire the original LED production images and the corresponding equipment operating data and environmental sensor data. The image processing module is used to perform super-resolution reconstruction processing on the original LED production image to obtain a high-resolution LED production image; A feature library construction module is used to construct a dynamic benchmark feature library for LED production based on preset LED production process standard data. A defect identification module is used to extract and identify defect features from high-resolution LED production images and generate a defect feature set. An error tracing module is used to trace the source of errors in the defect feature set and generate a corresponding error correction strategy based on the error tracing results. The feedback optimization module is used to perform feedback optimization on the dynamic benchmark feature library for LED production based on the implementation effect of the error correction strategy.

[0014] Compared with the prior art, the present invention has the following beneficial effects: Image acquisition is performed on each stage of LED production to obtain original production images of LEDs and corresponding equipment operating data and environmental sensor data, thereby integrating the scattered data from each stage of LED production and providing sufficient data input for subsequent data analysis. Super-resolution reconstruction is performed on the original LED production images to obtain high-resolution LED production images. Combined with preset LED production process standard data, a dynamic benchmark feature library for LED production is constructed. The super-resolution reconstruction process improves the clarity of the original LED production images, providing accurate data input for subsequent image analysis. The construction of the dynamic benchmark feature library for LED production also provides data support for subsequent defect detection and error tracing. Based on the LED production defect detection mechanism, defect features are extracted and identified from high-resolution LED production images to generate a defect feature set. High-resolution production images provide rich detailed information for defect feature extraction. Combined with the LED production defect detection mechanism, minute defects in the manufactured products can be identified. By generating a defect feature set, the description of defects is transformed from a qualitative judgment dominated by subjective experience to a quantitative analysis dominated by objective factors. Based on the error source tracing and correction mechanism, the defect feature set is used to trace the source of error, and the corresponding equipment operating data and environmental sensor data are combined to generate error correction strategies. The error source tracing mechanism can accurately locate the error source, reduce the investigation time, and generate targeted correction strategies to improve the improvement effect of error correction. The error correction strategy generates production parameter adjustment instructions and sends them to the corresponding LED production control equipment for closed-loop control. The implementation effect of the error correction strategy is evaluated by monitoring the status of the LED production control equipment after the correction strategy is executed. Based on the implementation effect, the LED production defect detection mechanism and error tracing correction mechanism are continuously optimized to achieve defect detection and error correction in LED production, thereby improving LED production quality and efficiency. Attached Figure Description

[0015] Figure 1 This is a flowchart of the steps of an LED production control method based on machine vision guidance provided by the present invention.

[0016] Figure 2 This is a schematic diagram of an LED production control system based on machine vision guidance provided by the present invention.

[0017] Figure 3 This is a flowchart of step S4 in a machine vision-guided LED production control method provided by the present invention. Detailed Implementation

[0018] The accompanying drawings in the embodiments provide a clear and complete description of the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can typically be arranged and designed in various different configurations.

[0019] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0020] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a schematic diagram of the execution flow of an LED production control method based on machine vision guidance, according to one embodiment of the present invention. Figure 3 This is a flowchart of step S4 in a machine vision-guided LED production control method provided by the present invention. The following is a detailed description of this machine vision-guided LED production control method.

[0021] Step S1: Image acquisition is performed on each stage of LED production to obtain original LED production images and corresponding equipment operating data and environmental sensor data.

[0022] In this embodiment, step S1 includes: Step S11: Image acquisition is performed on each stage of LED production.

[0023] Specifically, a multispectral camera array is used to synchronously acquire images of each stage of LED production, including die bonding, wire bonding, encapsulation, and aging testing, to obtain original production images of LEDs. The acquisition frequency is dynamically adjusted according to the production speed. In one possible embodiment, the multispectral camera array includes an ultraviolet camera, a visible light camera, and an infrared camera. The ultraviolet camera uses a deep ultraviolet band detector to capture hidden cracks caused by lattice defects inside the chip. The visible light camera uses a high-sensitivity CCD sensor to capture surface features such as the shape and position of the bonding wires. The infrared camera detects the temperature distribution differences in the bonding wire area to capture local overheating caused by poor contact. During image acquisition, a trigger synchronization mechanism is employed. The encoder on the production line sends a trigger signal to control the three types of cameras in the multispectral camera array to expose simultaneously, avoiding image misalignment caused by shooting time differences. At the same time, the image acquisition frequency is dynamically adjusted based on the real-time speed feedback of the LED production line. When the production speed increases, the camera exposure time is automatically shortened, and the frame rate is increased to ensure that each LED component is captured with at least 3 to 5 frames of images in each production stage, ensuring that no critical production moments are missed. For example, when the production speed increases from 5 LED components per second to 10 LED components per second, the image acquisition frequency will increase from 10 frames per second to 20 frames per second, and the exposure time will decrease from 10ms to 5ms. Furthermore, the insufficient light flux caused by the shortened exposure time is compensated for by increasing the sensor gain.

[0024] Step S12: Obtain the original LED production image and the corresponding equipment operating data.

[0025] Specifically, real-time operating data of the production equipment is collected synchronously through the equipment operating condition sensor group to obtain equipment operating condition data, which includes at least welding head temperature, welding pressure, die bonding accuracy, motor running speed and conveyor belt position.

[0026] Step S13: Obtain the original LED production image and the corresponding environmental sensor data.

[0027] Specifically, environmental sensor data corresponding to the LED production environment is collected through an environmental sensor array. The environmental sensor data includes at least ambient temperature, ambient humidity, ambient air pressure, air dust concentration, and light intensity.

[0028] Step S14: Perform data preprocessing on the original LED production images, equipment operating data, and environmental sensor data.

[0029] Specifically, the timestamps of the original LED production images, equipment operating data and environmental sensor data are uniformly calibrated, using the main clock of the LED production line as the reference, to eliminate drift errors caused by clock differences between various devices. By using linear interpolation, data from different sampling frequencies are adjusted to the same time interval. For the original LED production image, the image frame closest to the target timestamp is found and its index is recorded to achieve alignment of the original LED production image with equipment operating data and environmental sensing data in the time dimension. For spatial alignment, the data from the position sensors set on the LED production line conveyor belt are used as a reference to correlate the positions of LED components in the original LED production image with the physical positions of the production line, and establish a spatial coordinate transformation relationship to achieve alignment in the spatial dimension. For equipment operating data and environmental sensing data, the minimum-maximum normalization method is used to map the data to the [0, 1] interval; for the original LED production image, the RGB image is first converted to a grayscale image, then the histogram equalization method is used to enhance the contrast of the image, and then the pixel value is normalized to the [0, 1] interval. The data preprocessing process also includes outlier detection and handling. The 3σ criterion is used to identify outliers, and for detected outliers, the average value of neighboring data is used to replace them, preventing outliers from affecting subsequent analysis. Simultaneously, the preprocessed data is compressed and stored using a lossless compression algorithm to reduce data storage space and improve data transmission efficiency.

[0030] Step S2: Perform super-resolution reconstruction processing on the original LED production image to obtain a high-resolution LED production image, and construct a dynamic benchmark feature library for LED production by combining it with preset LED production process standard data.

[0031] In this embodiment, step S2 includes: Step S21: Perform super-resolution reconstruction processing on the original LED production image to obtain a high-resolution LED production image.

[0032] Specifically, a generative adversarial network with embedded spatial attention and channel attention mechanisms is used as the super-resolution reconstruction model, and the original LED production image is input into the super-resolution reconstruction model; After the original LED production image is input into the model, multi-scale feature extraction is performed through a seven-layer cascaded convolutional neural network. Simultaneously, feature enhancement and cross-layer fusion reconstruction are performed based on the residual dense network. Based on the adversarial training mechanism, and combined with preset structural parameters and loss function, iterative optimization is performed to generate high-resolution LED production images.

[0033] Understandably, a seven-layer cascaded convolutional neural network consists of a feature extraction layer, a feature enhancement layer, and an image reconstruction layer. The first three layers of the seven-layer cascaded convolutional neural network are feature extraction layers, which use 32, 64, and 128 3×3 convolutional kernels respectively, and extract features from the original LED production image through the ReLU activation function. In one possible embodiment, in the first feature extraction layer, the LeakyReLU activation function is selected to preserve gradient information of weak features such as oxide traces on the surface of the solder lines; in the second feature extraction layer, batch normalization of a 5×5 window is introduced, and the local correlation of gray-level fluctuations in the solder line region is utilized, such as the gray-level variance within adjacent 5×5 pixels being less than 10, to smooth noise while preserving real texture details; in the third feature extraction layer, for repetitive patterns such as periodically arranged electrodes on the chip solder joints, some neurons are randomly masked to prevent model overfitting.

[0034] The middle three layers of the seven-layer cascaded convolutional neural network are feature enhancement layers. Each feature enhancement layer can be represented as a residual dense block. Each residual dense block contains a 1×1 convolutional kernel and three 3×3 convolutional kernels. The feature maps output by the four convolutional kernels are concatenated through dense connections. A spatial attention module is set after each residual dense block. This module generates a spatial attention weight map by performing global average pooling and convolution operations on the feature map. The feature map is weighted based on the spatial attention weight map to highlight the features of the region of interest and generate the corresponding feature vector and defect feature label. In one possible embodiment, feature dimensionality reduction is performed using a 1×1 convolution kernel, with kernel parameters set using the He initialization method, thereby achieving cross-channel information fusion. A 3×3 convolution kernel is then used for detail capture. Taking the bonding region between the bonding wire metal and the electrode oxide layer as an example, assuming the feature map of channel 30 primarily captures the specular reflection effect on the bonding wire metal surface, this region exhibits a strong high-brightness response in the high-frequency signal band; while the feature map of channel 60 focuses on the dielectric properties of the electrode oxide layer, its signal distribution exhibiting a low-frequency grayscale gradient change. Cross-channel information fusion is performed using a 1×1 convolution kernel, and detail capture is performed using a 3×3 convolution kernel to obtain the electromagnetic property differences at the interface between the two materials. When a loose connection defect occurs, the transition region between the metal and the oxide layer will produce abnormal reflection coefficient fluctuations. This signal characteristic manifests as a sudden change in local grayscale values ​​on the convolutional feature map. Combined with the corresponding spatial attention weight map, feature weighting processing is performed to form an identifiable defect feature identifier (this is only an example illustrating the role of residual dense blocks; the specific content needs to be determined according to the actual situation).

[0035] The last layer of the seven-layer cascaded convolutional neural network is the image reconstruction layer. It uses a 3×3 convolutional kernel to map the feature vector output by the feature enhancement layer to a high-resolution image. For example, it maps a 128-dimensional feature vector to 3-dimensional RGB to generate a high-resolution production image of LEDs. Understandably, the embedding of the channel attention mechanism is represented as follows: after each convolutional layer, two one-dimensional feature vectors are obtained by performing global average pooling and global max pooling on the feature maps of each channel. These vectors are then input into a shared network containing two fully connected layers, which outputs channel attention weights and weights the feature maps of each channel to enhance the features of important channels. Understandably, during adversarial training, the loss function consists of three parts: perceptual loss, L1 pixel loss, and adversarial loss. The perceptual loss represents the feature difference between the generated high-resolution LED production image and the original LED production image. The L1 pixel loss represents the pixel value difference between the generated high-resolution LED production image and the original LED production image. The adversarial loss represents the difference in realism between the high-resolution LED production image and the original LED production image. In one possible embodiment, assuming super-resolution reconstruction is performed on the original 256×256 LED bonding wire image to generate a high-resolution LED production image G(x) with a resolution of 1024×1024, and the original LED production image is y, the third convolutional layer of the fourth convolutional block in the VGG19 network model is selected as the feature extractor to extract the high-resolution production image. The original production image of the LED is The feature maps are obtained by inputting them into the feature extractor. and The corresponding feature map size is 64×64×256. With a batch size of N=1, the loss for a single image is calculated based on the perceptual loss function. L1 loss function Adversarial loss function Calculate the corresponding image loss, where C, H, and W represent the number of channels, height, and width of the feature map, respectively, and (i,j,k) represents the pixel coordinates and channel index. Represented as a high-resolution production image output by a super-resolution reconstruction model. Given the probability of a real image, the corresponding perceptual loss, L1 pixel loss, and adversarial loss are 0.0122, 0.9537, and 0.16, respectively. They are weighted and summed according to a weight ratio of 5:3:2 to obtain the final image loss of 0.3242 (this is just an example of one possible form of the loss function; the specific form of the loss function needs to be determined according to the actual situation).

[0036] Step S22: Construct a dynamic benchmark feature library for LED production by combining preset LED production process standard data.

[0037] Specifically, the system analyzes the preset LED production process standard data, extracts the key quality parameters of each production stage in the LED production process, including die bonding, wire bonding, packaging, and aging testing, and generates the corresponding quality indicators for each production stage. In one possible embodiment, when parsing preset LED production process standard data, natural language processing technology is used to extract and classify keywords, perform part-of-speech tagging, and identify production entities in the process standard document. Key quality parameters and process requirements are extracted. For example, from a preset LED production process standard data in a certain process standard document, "Wire bonding process requirements: wire bonding tensile force not less than 6g, wire bonding diameter deviation controlled within ±0.5μm, contact area between wire bonding and electrode not less than 0.01mm²", three key quality parameters, wire bonding tensile force, wire bonding diameter deviation, and contact area, are extracted. The corresponding thresholds are determined according to the corresponding process requirements. The extracted key quality parameters are classified according to the LED production stages to generate a quality index system for each production stage. Each quality index includes at least the parameter name, unit, threshold range, and detection method.

[0038] Based on the quality indicators, feature extraction is performed on the high-resolution LED production images corresponding to each production stage in the qualified production batch to generate standard feature templates for each production stage. In one possible embodiment, when extracting features from high-resolution production images of qualified LED production batches based on quality indicators, a multi-scale feature extraction algorithm is used. For geometric dimension parameters, the contour of the target is extracted using an edge detection algorithm, and the contour is fitted using the least squares method to obtain corresponding contour data. The contour data includes at least the length, angle, and area corresponding to the contour. For gray-level distribution features, image gray-level features are extracted by calculating statistical data of the high-resolution LED production images. The statistical data includes at least the gray-level histogram, mean, and variance. For texture features, texture features are extracted using local binary mode and gray-level co-occurrence matrix. For example, in the wire bonding process, the geometric dimensions of the wire are extracted to obtain the required error for the wire length (within [-0.1, 0.1] μm), the required error for the wire diameter (within [-0.05, 0.05] μm), and the required error for the wire angle (within [-0.1, 0.1] °). The grayscale features of the wire bonding area are extracted to obtain the grayscale mean error of the image (within [-1, 1]). (This is just an example of the extracted feature form; the specific parameter data needs to be determined according to the actual situation.)

[0039] Based on the standard feature template, a corresponding defect feature sub-library is constructed for each production stage. Each defect feature sub-library includes a defect feature vector database, the process parameter range range and quality judgment threshold for the corresponding production stage, and assigns a time tag and batch information to each defect feature. The time tag and batch information are used to dynamically update the LED production dynamic benchmark feature library. In one possible embodiment, when constructing the defect feature sub-library, an object-oriented database design method is adopted. Each defect feature sub-library contains multiple data tables, which store defect feature vectors, process parameter ranges, and quality judgment thresholds, respectively. For the defect feature vector database, feature vectors are stored in the form of a vector space, and a unique index is established for each feature vector to improve query efficiency. The process parameter range is determined based on historical production data and process standards. For example, the welding pressure range is 3.5-4.5N. The process capability index is calculated using statistical process control methods, and the rationality of the process parameter range is verified based on the process capability index. The quality judgment threshold is determined based on quality indicators and the degree of impact of defects on product performance. For example, the threshold for a slight defect of a certain weld wire diameter deviation is [-0.3, 0.3] μm, the threshold for a moderate defect is [-0.5, 0.5] μm, and the threshold for a severe defect is [-0.7, 0.7] μm. In one possible embodiment, when dynamically updating the LED production dynamic benchmark feature library, an incremental learning algorithm is used. It is assumed that features are extracted from newly produced qualified products every 80 production batches to obtain real-time features. These real-time features are then compared with existing features in the LED production dynamic benchmark feature library, and the similarity between them is calculated. For real-time features with a similarity lower than a preset similarity threshold, they are added as new features to the LED production dynamic benchmark feature library. For real-time features with a similarity reaching the preset similarity threshold, the confidence level corresponding to the feature is updated, and outdated features that have not appeared in the library for a long time are deleted to maintain the timeliness and effectiveness of the feature library. For example, when producing a new model of LED, its bonding wire material and process change. Incremental learning is used to add the bonding wire features of the new model to the defect feature sub-library, enabling the feature library to adapt to new production needs, and deleting outdated features that have failed to match the real-time features three times consecutively.

[0040] Step S3: Based on the LED production defect detection mechanism, extract and identify defect features from the high-resolution LED production image to generate a defect feature set.

[0041] In this embodiment, step S3 includes: Step S31: Based on the LED production defect detection mechanism, perform defect feature recognition on the high-resolution LED production image to obtain candidate defect areas.

[0042] Specifically, a defect detection model is constructed by training a generative adversarial network based on federated learning and meta-learning. High-resolution production images of LEDs are input into the defect detection model to identify defects and generate corresponding defect probability heatmaps. In one possible embodiment, when constructing a defect detection model based on federated learning and meta-learning to train a generative adversarial network, the federated learning adopts a parameter server architecture, with multiple LED production sites as clients. Each client trains a sub-model of the generative adversarial network locally. Each LED production site uploads the parameters corresponding to the generated sub-model to the central server. The central server aggregates the sub-model parameters uploaded by each client, such as by using a weighted average method, where the weights are determined based on the amount of data from the client and the model performance. The aggregated sub-model parameters are then redistributed to each client for iterative training. Meta-learning employs a model-independent meta-learning algorithm. During training, different tasks are set for each client, such as different types of defect detection. By training on multiple tasks, the model parameters are given good initial values, enabling them to converge quickly on new tasks with a small number of samples.

[0043] Furthermore, the generator based on the defect detection model generates a defect probability heatmap with the same size as the input image. Each pixel value in the defect probability heatmap represents the probability that a defect exists at that location. The discriminator is used to distinguish the generated defect heatmap from the pre-set real defect annotation map. During training, the cross-entropy loss function and the adversarial loss function are used to enable the model to accurately generate the defect probability heatmap.

[0044] A non-maximum suppression algorithm with an adaptive threshold is used to filter defect regions in the defect probability heatmap to obtain candidate defect regions; In one possible embodiment, when using an adaptive threshold nonmaximum suppression algorithm to screen the defect probability heatmap, the initial threshold is first automatically determined based on the histogram distribution of the defect probability heatmap, such as taking the 90th percentile of the probability value as the initial threshold, and then the heatmap is binarized to obtain the defect candidate region. In the non-maximum suppression process, the confidence level of each defect candidate region is calculated, such as using the average probability value within the region as the confidence level. The defect candidate regions are then sorted according to their confidence levels. Starting with the defect candidate region with the highest confidence level, other regions with an overlap greater than a preset overlap threshold are suppressed to obtain the final defect candidate regions. For example, the intersection-union ratio (IUGR) can be used as the region overlap. For other regions with an overlap greater than 0.5, regions with a confidence level of at least 0.96 are retained and used as the final candidate defect regions. For instance, in the defect probability heatmap of a weld wire, after adaptive thresholding and non-maximum suppression, three candidate defect regions are selected with confidence levels of 0.98, 0.97, and 0.96, respectively. These regions can all be considered as potentially defective areas, so they are output as candidate defect regions.

[0045] Step S32: Extract defect features from candidate defect regions to generate a defect feature set.

[0046] Specifically, defect features are extracted from candidate defect regions to generate defect feature vectors, and the defect feature vectors are compared with the corresponding defect feature vector database in the LED production dynamic benchmark feature library. In one possible embodiment, when extracting defect features from candidate defect regions, the same feature acquisition method used when constructing the dynamic benchmark feature library for LED production is employed to extract defect features from candidate defect regions, and a feature vector of the defect to be detected is constructed based on the extracted defect features. The feature vector of the defect to be detected is compared with the defect feature vector database contained in the LED production dynamic benchmark feature library. The k-nearest neighbor algorithm is used to find the k most similar samples to the feature vector of the defect to be detected. Based on the information corresponding to these samples, the defect type is initially determined. The feature vector of the defect to be detected is then input into the classifier of the support vector machine algorithm to further determine the preliminary defect type and obtain the final defect comparison result. The support vector machine algorithm classifier is trained using sample data in the LED production dynamic benchmark feature library. Furthermore, based on the comparison results, the defect type and severity level of the candidate defect area are labeled, and the spatial coordinate information corresponding to the defect in the high-resolution LED production image is recorded simultaneously. The defect type, severity level and spatial coordinate information are then encapsulated to generate a defect feature set.

[0047] In one possible embodiment, assuming that in the three candidate defect regions of the wire bonding process, each region is cropped and defect features are extracted, and the defect feature vector of one region is matched with the defect feature vector database in the LED production dynamic benchmark feature library. If four out of the five samples matched by the k-nearest neighbor algorithm are loose wire connections, then the defect feature vector with the loose wire connection is determined to have the highest similarity. Furthermore, the vector machine algorithm classifier also further determines that the defect is a loose wire connection, so the fault type is determined to be a loose wire connection. The wire diameter deviation of the candidate defect region is obtained to be 0.4 μm, which belongs to a moderate defect. Its spatial coordinate information in the high-resolution LED production image is obtained, and the corresponding information is encapsulated to generate the corresponding defect feature set.

[0048] Step S4: Based on the error source tracing and correction mechanism, the defect feature set is traced for error source, and an error correction strategy is generated by combining the corresponding equipment operating condition data and environmental sensing data.

[0049] In this embodiment, step S4 includes: Step S41: Based on the error source correction mechanism, perform error source tracing on the defect feature set to obtain the error source.

[0050] In this embodiment, step S41 includes: Step S41-1: Obtain the defect types in the LED production process and associate them with the corresponding equipment operating data and environmental sensor data.

[0051] Specifically, based on the LED production dynamic benchmark feature library, the defect types corresponding to defects generated during the LED production process are obtained, and the equipment operating condition data and environmental sensor data with the highest correlation are preset for different defect types.

[0052] In one possible embodiment, when pre-setting the equipment operating condition data and environmental sensor data with the highest correlation for different defect types, the defect types corresponding to defects generated during LED production are obtained based on the LED production dynamic benchmark feature library, and a defect parameter correlation matrix is ​​constructed by combining historical production data and expert experience. The elements in the matrix represent the correlation degree between a certain defect type and a certain parameter. The correlation degree is obtained by calculating the mutual information algorithm, and the larger the mutual information value, the closer the correlation between the parameter and the defect type. For example, the correlation degree between the wire bonding defect and the welding head temperature, welding pressure and ambient humidity is 0.8, 0.7 and 0.6, respectively, while the correlation degree with other parameters is less than 0.3. Therefore, the welding head temperature, welding pressure and ambient humidity are preset as the main correlation parameters for the wire bonding defect.

[0053] Step S41-2: Construct an error source tracing decision tree with the defect type as the root node, the corresponding equipment operating data and environmental sensor data as intermediate nodes, and the error source as the leaf node.

[0054] Specifically, the C4.5 decision tree algorithm is used, with the defect type as the root node. The parameter with the highest correlation is selected as the first splitting attribute. Branching is performed according to the value range of this parameter. For each branch node, the correlation between the remaining parameters under that branch node and the defect type is calculated. The parameter with the highest correlation is selected as the new splitting attribute for the next level of branching, until the preset stopping conditions are met, such as the number of samples in the node is less than a threshold, all samples belong to the same category, or there are no more available splitting attributes. Finally, a complete error source tracing decision tree is formed. Each internal node of the decision tree contains an attribute name, a splitting threshold, and an information gain value. The information gain value is used to measure the degree of contribution of the attribute to defect classification.

[0055] In one possible embodiment, taking a poorly connected wire bonding defect as an example, the welding head temperature is first used as the first splitting attribute, and the temperature range is divided into three branches: below 220℃, [220, 240]℃, and above 240℃. Each branch is further divided with welding pressure as the splitting attribute, and so on, until the leaf node is a specific error source. For example, welding head temperature that is too low, welding pressure that is insufficient, and ambient humidity that is too high can all be defined as specific error sources (this is just an example of the construction process of the error source tracing decision tree, and the specific content needs to be determined according to the actual situation).

[0056] Step S41-3: Combine equipment operating data and environmental sensor data to trace the source of error in the defect feature set.

[0057] Specifically, the defect feature set and the corresponding equipment operating condition data and environmental sensor data are input into the error tracing decision tree. Based on the defect type contained in the defect feature set, the system matches and locates the corresponding branch node in the error tracing decision tree. The system then adjusts the equipment operating condition data and environmental sensor data corresponding to the defect type, analyzes the changes in the defect features corresponding to the defect type, and locates the error source corresponding to the defect feature set.

[0058] In one possible embodiment, when inputting the defect feature set and the corresponding equipment operating condition data and environmental sensing data into the error tracing decision tree, the root node corresponding to the error tracing decision tree is first located according to the defect type of the defect feature set. Then, according to the branching logic of the error tracing decision tree, the actual values ​​of the equipment and environmental parameters, the equipment operating condition data and the environmental sensing data are compared with the splitting threshold corresponding to the root node in turn, and the corresponding branches are traversed downwards until the leaf node is reached, thereby determining the error source. During the traversal, when the values ​​of equipment operating data and environmental sensor data deviate from the corresponding preset standard values, the deviation between the actual value and the set standard value is calculated and recorded in the error tracing process. Taking the error tracing of a loose weld wire defect as an example, the actual value of the welding head temperature is 227℃, while the preset standard value of the welding head temperature is 230℃, so the corresponding welding head temperature deviation is -3℃; the actual value of the welding pressure is 3.8N, while the corresponding standard value of the welding pressure is 4.0N, so the corresponding welding pressure deviation is -0.2N; the actual value of the ambient humidity is 55%, while the corresponding standard value of the ambient humidity is 50%, so the corresponding ambient humidity deviation is +5%. Through decision tree traversal, the error sources are finally located as low welding head temperature and insufficient welding pressure.

[0059] Step S42: Based on the error source, and combined with equipment operating data and environmental sensor data, an error correction strategy is generated.

[0060] Specifically, an error source tracing report is generated based on the error source. An LED production state space is constructed based on the defect feature set, the error source tracing report, equipment operating data, and environmental sensor data. Error correction is performed based on reinforcement learning.

[0061] In one possible embodiment, when generating an error tracing report, the error source type, its impact on product quality, associated equipment operating data and environmental sensor data, and corresponding deviation values ​​are recorded in detail according to a preset template format. The error source type is categorized into abnormal equipment parameters, abnormal environmental parameters, and abnormal material parameters. The impact level is assessed based on the corresponding severity level and its effect on the manufactured product. For example, a moderate wire bonding defect may lead to unstable electrical performance of LED components and a 30% reduction in lifespan. The associated equipment operating data, environmental sensor data, and deviation values ​​are directly referenced from the data recorded during the error tracing process. For instance, a partial error tracing report for a wire bonding defect might look like this: {Error source type: Abnormal equipment parameters; Impact level: Moderate, may lead to unstable electrical performance and a 30% reduction in lifespan; Associated parameters and deviation values: Solder head temperature 227℃ (standard 230℃, deviation -3℃), welding pressure 3.8N (standard 4.0N, deviation -0.2N)}.

[0062] Furthermore, the reinforcement learning agent, based on a preset reward function, continuously interacts and learns with equipment operating data and environmental sensor data within the LED production state space to generate targeted error correction strategies.

[0063] Specifically, the reinforcement learning agent is trained by continuously interacting with the production environment. At each time step, the reinforcement learning agent monitors the current production state and selects an action based on the probability distribution output by the reinforcement learning model. After the action is executed, the environmental state changes accordingly, and the reinforcement learning agent receives a reward according to a preset reward function. Through continuous iterative training, the reinforcement learning agent learns the optimal correction action for errors under different states.

[0064] In one possible embodiment, when constructing the LED production state space, the defect feature set, error tracing report, and equipment environmental data, equipment operating condition data, and environmental sensor data are used as state variables, and a preset value range is set for each state variable, such as the value range of the welding head temperature is defined in [200, 300]℃, and the value range of the welding pressure is defined in [3, 5]N; The reward function for reinforcement learning is set as R = 0.6 × defect elimination rate + 0.2 × production efficiency improvement rate - 0.2 × parameter adjustment cost, where the defect elimination rate is the ratio of the number of defects before and after the correction, the production efficiency improvement rate is the ratio of the output per unit time before and after the correction, and the parameter adjustment cost is positively correlated with the adjustment magnitude and frequency. For example, after a certain adjustment, if the defect elimination rate is 80%, the production efficiency improvement rate is 5%, and the parameter adjustment cost is 10%, then R = 0.6 × 0.8 + 0.2 × 0.05 - 0.2 × 0.1 = 0.47.

[0065] Step S5: Generate production parameter adjustment instructions according to the error correction strategy and send them to the corresponding LED production control equipment for closed-loop control.

[0066] In this embodiment, step S5 includes: Step S51: Generate production parameter adjustment instructions based on the error correction strategy.

[0067] In this embodiment, step S51 includes: Step S51-1: Convert the error correction strategy into digital control instructions that can be directly executed by the corresponding LED production equipment.

[0068] Specifically, the error correction strategy is converted into digital control instructions that can be directly executed by the corresponding LED production equipment, and the timestamps of the digital control instructions are synchronized using a time-sensitive network-based timestamp synchronization mechanism.

[0069] Step S51-2: Verify the rationality of digital control instructions based on the instruction rationality verification mechanism.

[0070] In one possible embodiment, the instruction rationality verification mechanism is carried out through a digital twin model. Digital control instructions are input into the LED production digital twin for production simulation. During the simulation, the focus is on whether parameter adjustments will lead to other quality problems, such as excessively high welding head temperature causing the wire bonding material to melt, and whether it is within the safe operating range of the equipment, such as welding pressure not exceeding the maximum bearing pressure of the equipment (5N). If the simulation results show that the defect elimination rate is greater than 99% and there are no other abnormalities, the instruction is deemed reasonable. For digital control instructions that pass the reasonableness verification, the digital control instruction is directly sent to the corresponding LED production control equipment. For digital control instructions that fail the rationality verification, return to the error correction strategy generation stage for modification and adjustment, and re-verify the rationality of the digital control instructions corresponding to the modified error correction strategy. For example, if the instruction to increase the welding head temperature by 3°C and the welding pressure by 0.2N is input into the digital twin model, and the simulation results show that the defect elimination rate of the weld wire connection is 99.5% and the equipment is operating normally, then the instruction is determined to have passed the rationality verification (this is just an example of one possible form of rationality verification mechanism; the specific content of the rationality verification mechanism needs to be determined according to the actual situation).

[0071] The system monitors the changes in the production image corresponding to the execution of digital control commands by the LED production control equipment in real time, evaluates the effects of the changes in the production image, and generates a closed-loop control report for LED production that includes the command content, execution process, and correction effects.

[0072] Step S52: Perform feedback optimization based on the LED production closed-loop control report.

[0073] Specifically, when the effect of N consecutive corrections is qualified, the current model parameters and production parameters are saved as optimization templates. Based on the production batch and product type corresponding to LED production, the value of N is set differently. For example, for regular batch production with a production quantity of more than 1,000 units per batch, N is set to 60 to ensure production stability; for small batch trial production with a production quantity of less than 1,000 units per batch, N is set to 30 to accelerate the debugging of new products; when no serious quality defects are found in the results of N consecutive tests, and the rate of minor quality defects is lower than the preset rate of minor defects, it is judged as a qualified test sequence.

[0074] Understandably, serious quality defects are defined as functional defects that directly affect the product's function and lifespan, while minor quality defects are defined as surface defects that do not affect the product's function and lifespan.

[0075] Furthermore, an optimization template is generated based on the tuning parameters and production parameters corresponding to the qualified inspection sequence. This optimized template is then matched with the standard feature templates corresponding to each defect feature sub-library in the LED production dynamic benchmark feature library. The tuning parameters represent the parameters used during super-resolution reconstruction processing, defect feature extraction and identification, and error tracing. The production parameters represent the optimal combination of equipment operating data and environmental sensor data corresponding to the qualified inspection sequence. For example, in the wire bonding stage of conventional batch production, after error correction, if 60 consecutive inspections show no serious defects and the minor defect rates are 0.07%, 0.06%,... all below the preset minor defect rate threshold of 0.08%, it is determined to be a qualified sequence. The corresponding tuning parameters are saved: {weights of the super-resolution reconstruction model, weights of the defect detection model} and production parameters: {welding head temperature: 230℃, welding pressure: 4.0N, ambient humidity: 50%}. The tuning parameters and production parameters are then used to generate an optimization template according to a preset standard template format (this is just an example of the optimization template generation process; specific data needs to be determined based on the actual content).

[0076] In one possible embodiment, when matching the optimized template with the dynamic benchmark feature library, the similarity between the features corresponding to the optimized template and the standard feature templates corresponding to each defect feature sub-library in the LED production dynamic benchmark feature library is calculated, such as by using a cosine similarity algorithm to calculate the similarity. If the maximum similarity is not less than 0.9, the match is considered successful. The standard feature template is updated based on the parameters in the optimized template, such as adjusting the baseline value and threshold range of the wire diameter. If the maximum similarity is less than 0.9, the matching is deemed to have failed. A new defect feature sub-library is constructed based on the optimized template, and the structure and content of the sub-library are consistent with those of the other sub-libraries.

[0077] Figure 2 The diagram illustrates a machine vision-guided LED production control system that can implement the ideas of this application, according to some embodiments of this application.

[0078] Specifically, a machine vision-guided LED production control system includes: The data acquisition module is used to acquire original LED production images and corresponding equipment operating data and environmental sensor data.

[0079] An image processing module is used to perform super-resolution reconstruction processing on the original LED production image to obtain a high-resolution LED production image.

[0080] The feature library construction module is used to construct a dynamic benchmark feature library for LED production based on preset LED production process standard data.

[0081] The defect identification module is used to extract and identify defect features from high-resolution LED production images and generate a defect feature set.

[0082] An error tracing module is used to trace the source of errors in the defect feature set and generate a corresponding error correction strategy based on the error tracing results.

[0083] The feedback optimization module is used to perform feedback optimization on the dynamic benchmark feature library for LED production based on the implementation effect of the error correction strategy.

[0084] The specific usage and function of this embodiment are explained below: Image acquisition is performed on each stage of LED production to obtain original production images of LEDs and corresponding equipment operating data and environmental sensor data, thereby integrating the scattered data of each stage of LED production and providing a complete and coherent data foundation for subsequent full-process data analysis. Super-resolution reconstruction is performed on the original LED production images to obtain high-resolution LED production images. Combined with preset LED production process standard data, a dynamic benchmark feature library for LED production is constructed. The super-resolution reconstruction process improves the clarity of the original LED production images, providing accurate data input for subsequent image analysis. Furthermore, the construction of the dynamic benchmark feature library for LED production provides standardized data support for subsequent defect detection and error tracing. Based on the LED production defect detection mechanism, defect features are extracted and identified from high-resolution LED production images to generate a defect feature set. High-resolution production images provide rich detailed information for defect feature extraction. Combined with the LED production defect detection mechanism, minute defects in the manufactured products can be identified. By generating a defect feature set, the description of defects is transformed from a qualitative judgment dominated by subjective experience to a quantitative analysis dominated by objective factors. The error source tracing and correction mechanism is used to trace the source of errors in the defect feature set. Combined with the corresponding equipment operating data and environmental sensor data, error correction strategies are generated. The error source tracing mechanism can accurately locate the error source, reduce the investigation time, and the generated error correction strategy is highly targeted and can accurately match the characteristics of the error source, thereby improving the effectiveness and efficiency of error correction. The error correction strategy generates production parameter adjustment instructions and sends them to the corresponding LED production control equipment for closed-loop control. The implementation effect of the error correction strategy is evaluated by monitoring the status of the LED production control equipment after the correction strategy is executed. Based on the implementation effect, the LED production defect detection mechanism and error tracing correction mechanism are continuously optimized to achieve defect detection and error correction in LED production, thereby improving LED production quality and efficiency.

[0085] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0086] It should be understood that, in the embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0087] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A machine vision-guided LED production control method, characterized in that, It includes the following steps: Image acquisition is performed on each stage of LED production to obtain original LED production images and corresponding equipment operating condition data and environmental sensor data. Super-resolution reconstruction processing is performed on the original LED production images to obtain high-resolution LED production images, and a dynamic benchmark feature library for LED production is constructed by combining it with preset LED production process standard data. Based on the LED production defect detection mechanism, defect features are extracted and identified from high-resolution LED production images to generate a defect feature set. Based on the error source tracing and correction mechanism, the defect feature set is used to trace the source of error, and error correction strategy is generated by combining the corresponding equipment operating condition data and environmental sensing data. The error correction strategy generates production parameter adjustment instructions, which are then sent to the corresponding LED production control equipment for closed-loop control.

2. The LED production control method based on machine vision guidance according to claim 1, characterized in that, Super-resolution reconstruction processing is performed on the original LED production images to obtain high-resolution LED production images, including: A generative adversarial network with embedded spatial attention and channel attention mechanisms is used as the super-resolution reconstruction model, and the original LED production image is input into the super-resolution reconstruction model. After the original LED production image is input into the model, multi-scale feature extraction is performed through a seven-layer cascaded convolutional neural network. Simultaneously, feature enhancement and cross-layer fusion reconstruction are performed based on the residual dense network. Based on the adversarial training mechanism, and combined with preset structural parameters and loss function, iterative optimization is performed to generate high-resolution LED production images.

3. The LED production control method based on machine vision guidance according to claim 1, characterized in that, A dynamic benchmark feature library for LED production is constructed by combining preset LED production process standard data, including: Analyze the preset LED production process standard data, extract the key quality parameters of each production link in the LED production process, such as die bonding, wire bonding, packaging and aging test, and generate the corresponding quality indicators for each production link. Based on the quality indicators, feature extraction is performed on the high-resolution LED production images corresponding to each production stage in the qualified production batch to generate standard feature templates for each production stage. Based on the standard feature template, a corresponding defect feature sub-library is constructed for each production stage. Each defect feature sub-library contains a defect feature vector database, the process parameter range range and quality judgment threshold for the corresponding production stage, and assigns a time tag and batch information to each defect feature. Among them, time tags and batch information are used to dynamically update the dynamic benchmark feature library for LED production.

4. The LED production control method based on machine vision guidance according to claim 3, characterized in that, Based on the LED production defect detection mechanism, defect features are extracted and identified from high-resolution LED production images to generate a defect feature set, including: A defect detection model is constructed by training a generative adversarial network based on federated learning and meta-learning. High-resolution LED production images are input into the defect detection model to identify defects and generate corresponding defect probability heatmaps. A non-maximum suppression algorithm with an adaptive threshold is used to filter defect regions in the defect probability heatmap to obtain candidate defect regions; Defect features are extracted from candidate defect regions to generate defect feature vectors, and the defect feature vectors are compared with the corresponding defect feature vector database in the LED production dynamic benchmark feature library. Based on the comparison results, the defect type and severity level of the candidate defect area are labeled, and the spatial coordinate information of the defect in the high-resolution LED production image is recorded simultaneously. The defect type, severity level and spatial coordinate information are then encapsulated to generate a defect feature set.

5. The LED production control method based on machine vision guidance according to claim 1, characterized in that, The error source tracing and correction mechanism is used to trace the source of errors in the defect feature set, and an error correction strategy is generated by combining the corresponding equipment operating data and environmental sensor data, including: Based on the LED production dynamic benchmark feature library, the defect types corresponding to defects generated during the LED production process are obtained, and the equipment operating condition data and environmental sensor data with the highest correlation are preset for different defect types. With the defect type as the root node, the corresponding equipment operating data and environmental sensor data as intermediate nodes, and the error source as the leaf node, an error source tracing decision tree is constructed. The defect feature set and the corresponding equipment operating condition data and environmental sensing data are input into the error tracing decision tree. Based on the defect type contained in the defect feature set, the tree is matched and located to the corresponding branch node at the end of the error tracing decision tree. The equipment operating condition data and environmental sensing data corresponding to the defect type are adjusted, the changes in the defect features corresponding to the defect type are analyzed, and the error source corresponding to the defect feature set is located. An initial error correction decision is generated based on the error source, and an error correction strategy is generated by combining the corresponding equipment operating data and environmental sensor data.

6. The LED production control method based on machine vision guidance according to claim 5, characterized in that, An initial error correction decision is generated based on the error source, and an error correction strategy is generated by combining the corresponding equipment operating data and environmental sensor data, including: An error source tracing report is generated based on the error source. An LED production state space is constructed based on the defect feature set, the error source tracing report, equipment operating data, and environmental sensor data. Error correction is performed based on reinforcement learning. The error tracing report should include at least the type of error source, the degree of impact on product quality, the associated equipment operating data and environmental sensor data, and the corresponding deviation value. The reinforcement learning agent, based on a preset reward function, learns by continuously interacting with equipment operating data and environmental sensor data within the LED production state space, generating targeted error correction strategies.

7. The LED production control method based on machine vision guidance according to claim 1, characterized in that, Based on the error correction strategy, production parameter adjustment instructions are generated and sent to the corresponding LED production control equipment for closed-loop control, including: The error correction strategy is converted into digital control instructions that can be directly executed by the corresponding LED production equipment, and the timestamps of the digital control instructions are synchronized using a time-sensitive network-based timestamp synchronization mechanism. Before the instruction is sent, the digital control instruction is validated for its rationality based on the instruction rationality verification mechanism; For digital control commands that have passed the rationality verification, the digital control command is directly sent to the corresponding LED production control equipment. For digital control instructions that fail the rationality verification, return to the error correction strategy generation stage for modification and adjustment, and re-verify the rationality of the digital control instructions corresponding to the modified error correction strategy. Real-time monitoring of changes in production images corresponding to the execution of digital control commands by LED production control equipment, evaluation of the effects of the changes in production images, and generation of a closed-loop control report for LED production that includes command content, execution process, and correction effects; When the effect of N consecutive corrections is satisfactory, save the current model parameters and production parameters as an optimization template.

8. The LED production control method based on machine vision guidance according to claim 1, characterized in that, When the effect of N consecutive corrections is satisfactory, save the current corresponding tuning parameters and production parameters as an optimization template, including: The N value is set differently based on the production batch and product type corresponding to LED production. When no serious quality defects are found in N consecutive test results, and the rate of minor quality defects is lower than the preset rate of minor defects, the test sequence is judged to be qualified. Among them, serious quality defects are functional defects that directly affect the product's function and service life, while minor quality defects are surface defects that do not affect the product's function and service life. An optimization template is generated based on the tuning parameters and production parameters corresponding to the qualified inspection sequence, and the optimization template is matched with the standard feature templates corresponding to each defect feature sub-library in the LED production dynamic benchmark feature library; If a standard feature template is successfully matched, the standard feature template is then optimized based on the optimized template. If no standard feature template is matched, a defect feature sub-library is constructed based on the optimized template; Among them, the tuning parameters refer to the parameters corresponding to the super-resolution reconstruction processing, defect feature extraction and identification, and error tracing operations, while the production parameters refer to the optimal combination of equipment operating data and environmental sensor data corresponding to the qualified inspection sequence.

9. The LED production control method based on machine vision guidance according to claim 1, characterized in that, Image acquisition is performed at each stage of LED production to obtain raw production images of LEDs and corresponding equipment operating data and environmental sensor data, including: The LED production process is synchronously captured using a multispectral camera array to obtain raw LED production images. The acquisition frequency is dynamically adjusted according to the production speed. Real-time operating data of the production equipment is collected synchronously by the equipment operating condition sensor group. The equipment operating condition data includes at least the welding head temperature, welding pressure, die bonding accuracy, motor running speed and conveyor belt position. The environmental sensor group collects environmental sensing data corresponding to the LED production environment. The environmental sensing data includes at least the ambient temperature, ambient humidity, ambient air pressure, air dust concentration, and light intensity. Data preprocessing is performed on raw LED production images, equipment operating data, and environmental sensor data. The data preprocessing includes at least spatiotemporal alignment and normalization operations.

10. A machine vision-guided LED production control system, characterized in that, include: The data acquisition module is used to acquire the original LED production images and the corresponding equipment operating data and environmental sensor data. The image processing module is used to perform super-resolution reconstruction processing on the original LED production image to obtain a high-resolution LED production image; A feature library construction module is used to construct a dynamic benchmark feature library for LED production based on preset LED production process standard data. A defect identification module is used to extract and identify defect features from high-resolution LED production images and generate a defect feature set. An error tracing module is used to trace the source of errors in the defect feature set and generate a corresponding error correction strategy based on the error tracing results. The feedback optimization module is used to perform feedback optimization on the dynamic benchmark feature library for LED production based on the implementation effect of the error correction strategy.

Images