EL battery piece defect intelligent classification method and system based on multi-model collaborative reasoning

Abstract

The invention discloses an EL battery piece defect intelligent classification method and system based on multi-model collaborative reasoning. The method comprises the steps of obtaining image data of a to-be-detected EL battery piece; three-stage collaborative reasoning processing is carried out on the image data, and the three-stage collaborative reasoning processing comprises the steps that first-stage coarse-grained screening is executed, preliminary defect classification is carried out on the image data through a parallel double-model architecture, and a first-stage classification result is obtained; second-level positioning analysis is carried out based on the first-level classification result, and when the first-level classification result is a preset specific category, a positioning model is triggered to carry out space positioning on the defect; third-level fine granularity judgment is executed, and when the first-level classification result is a preset subdivision category, a special subclass judgment model is called to obtain a second-level subdivision result; and outputting the primary classification result and / or the secondary classification result, and performing association analysis according to the classification result and a preset defect-process database. The method is high in classification accuracy, high in processing speed and intelligent in process tracing.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of photovoltaic module quality detection, and particularly relates to an EL cell defect intelligent classification method and system based on multi-model collaborative reasoning. BACKGROUND

[0002] In the production and manufacturing process of photovoltaic modules, quality detection is a key link to ensure its performance and reliability. Electroluminescence (EL) detection technology has become the mainstream quality control method in the industry because it can directly reveal the internal microscopic defects of the cell. In the early manual inspection stage, the detection efficiency is low and highly dependent on human experience, and only obvious defects can be identified, which has the problems of subjectivity and missed detection. Then in the traditional image processing stage, edge detection, morphological processing and other methods are used to realize partial automation. Although this stage improves the efficiency, it has a high misjudgment rate for complex defects (such as cloud-shaped marks) or defects with similar visual features (such as furnace belt marks and needle marks), and the algorithm parameters need to be frequently adjusted by humans, resulting in poor adaptability and high maintenance cost. The introduction of deep learning technology has significantly improved the automation level and accuracy of EL defect detection. Models represented by convolutional neural networks (CNN) realize end-to-end classification, and object detection algorithms (such as the YOLO series) also begin to be applied to defect positioning; it also includes the use of Transformer models, which have an average precision mean (mAP) of 89.7% on EL datasets.

[0003] However, the existing deep learning solutions have the following problems in industrial large-scale deployment: a single classification model (such as ResNet50) has a low classification accuracy when dealing with highly similar visual feature pairs of defects (such as "black edge" and "quartz boat mark", "black corner" and "boat tooth black edge"), which cannot meet the requirements of refined quality grading. Existing solutions usually only output one final defect label, which cannot perform deeper analysis. For example, after detecting "black edge", it cannot further distinguish whether it is caused by boat tooth obstruction or process pollution, which limits the ability to trace the cause of the defect. Photovoltaic production lines have very high detection speed requirements (throughput demand greater than 1000 pieces / minute), and complex models with high precision often have slow reasoning speed.

[0004] Therefore, how to design an EL cell defect intelligent classification scheme that can accurately distinguish between confusing defects, provide multi-level classification information, balance high precision and high efficiency, and be associated with production processes, is a technical problem that needs to be solved in the current field. SUMMARY

[0005] The present application aims to provide an EL cell defect intelligent classification system and method based on multi-model collaborative reasoning, which mainly solves the following technical problems: Due to the dependence on manual feature engineering and fixed parameters, the detection rate of micro-cracks and low-contrast defects is high, and the production process cannot be adapted to the changing production process. The single deep learning model has low classification accuracy when dealing with confused defects with similar visual features, and cannot distinguish between defect subclasses (such as "chaotic black corner" and "boat tooth black edge"). Solve the problem that the standard classification model cannot output the defect position coordinates, making it difficult to accurately position and measure the size. Make up for the shortcomings of existing detection technology, which only outputs isolated defect results and cannot be associated with upstream production process parameters for analysis, making it difficult to achieve quality closed-loop tracking and optimization. To achieve the above purpose, an embodiment of the present application provides a kind of EL cell piece defect intelligent classification method based on multi-model collaborative reasoning, the method comprises: Obtain the image data of the EL cell piece to be detected; Three-level collaborative reasoning processing is carried out on the image data, which includes: performing a primary coarse-grained screening, performing a preliminary defect classification on the image data through a parallel double-model architecture, and obtaining a primary classification result;Based on the primary classification result, the positioning model is triggered when the primary classification result is a preset specific category, and the spatial positioning of the defect is carried out;Performing a three-level fine-grained discrimination, when the primary classification result is a preset fine classification category, a special subclass discrimination model is called to obtain a secondary subdivision result; Output the primary classification result and / or the secondary subdivision result at the application layer, and perform correlation analysis according to the classification result and the preset defect-process database.

[0006] In one or more embodiments of the present application, the parallel double-model architecture comprises: The first model is used to detect high-priority hidden cracks or hidden crack defects and output their position coordinates; The second model is used to classify the remaining preset defect categories; Wherein, when the first model detects a hidden crack defect, the result is determined first and the subsequent reasoning process can be interrupted.

[0007] In one or more embodiments of the present application, the preset fine classification category includes "EL black edge_black corner" and "EL dirt". When the primary classification result is "EL black edge_black corner", the black edge subdivision model is called to distinguish it into one of boat tooth black edge, black edge_black corner up and down, and black edge_black corner chaos; When the primary classification result is "EL dirt", the dirt subdivision model is called to distinguish it into one of dirt dark film, dirt plating, dirt backflow, oxidation dirt and dirt chaos.

[0008] In one or more embodiments of the present application, the second model comprises an Inception_v3 classification model and a router connected to the Inception_v3 classification model, wherein the router integrates a dynamic routing mechanism and a double-branch structure; The dynamic routing mechanism generates a category weight vector according to the image features extracted from the backbone network through a routing controller, and dynamically assigns the features to different processing branches. The double-branch structure comprises a regular defect classification branch and a dirty defect classification branch.

[0009] In one or more embodiments of the present application, the dynamic routing mechanism further comprises a learnable weight adjustment parameter normal_boost for dynamically adjusting the weight proportion of the regular defect category; and the gradient of the dirty defect classification branch is subjected to decay processing during model training to suppress feature interference and prevent overfitting.

[0010] In one or more embodiments of the present application, the adjustment formula of the weight of the regular defect category is: ; ; wherein, x is a 2048-dimensional feature vector output by the backbone network, representing multi-scale defect features of the battery piece EL image; , is a weight matrix for linear transformation; ∈ is used to compress high-dimensional features to 512 dimensions to extract key defect features; ∈ is used to generate a category weight vector corresponding to different defect types; b 1、 b 2 is a bias term to enhance the flexibility of the model; Hardswish is an activation function that balances the non-linear expression ability and computational efficiency, and is used to process complex defect patterns in the EL image.

[0011] σ represents a Sigmoid function that normalizes the weight to [0, 1], indicating the priority of each defect category; normal boost is an initial learnable parameter used to dynamically boost the weight of the regular defect.

[0012] In one or more embodiments of the present application, the conventional defect classification branch and the contamination defect classification branch adjust logits of model output using different temperature parameters τ to control the sharpness of the Softmax probability distribution, wherein the temperature parameter of the conventional defect classification branch is lower than that of the contamination defect classification branch; wherein the formula for controlling the sharpness of the Softmax probability distribution of logits using the temperature parameter τ includes: ; wherein logits are original output scores of the model for the i-th class of defects, used to reflect the confidence of the i-th class of defects in the EL image; i τ is the temperature parameter, wherein in the conventional branch τ = 0.05; in the contamination branch τ = 0.8, a higher temperature smooths the probability distribution to tolerate low-confidence contamination samples to improve fault tolerance; P(y = i) is the final classification probability, and the sharpness thereof directly affects the strictness of EL detection.

[0013] In another aspect of the present application, an EL cell piece defect intelligent classification system based on multi-model collaborative reasoning is provided, which comprises: a device layer comprising at least one data acquisition terminal for acquiring EL cell piece image data to be detected; a computing power model layer connected to the device layer, configured to receive and process the image data, and comprising the following inside: a first-level coarse-grained screening module configured to perform preliminary defect classification on the image data through a parallel double-model architecture to obtain a first-level classification result; a second-level positioning analysis module configured to perform spatial positioning of defects when the first-level classification result is a preset specific class; a third-level fine-grained discrimination module configured to call a dedicated sub-class discrimination model when the first-level classification result is a preset sub-classifiable class to obtain a second-level subdivision result; an application layer connected to the computing power model layer, configured to output the first-level classification result and / or the second-level subdivision result, and provide a human-computer interaction interface.

[0014] In one or more embodiments of the present application, the double-model architecture in the first-level coarse-grained screening module comprises: a first model unit, which can use a YOLOv8 target detection model, for detecting high-priority hidden cracks or crack-like defects; ​The second model unit can adopt a DDB-Inception classification model for classifying the remaining preset defect categories, and the DDB-Inception classification model comprises an Inception_v3 backbone network, a dynamic routing controller, and a general defect classification branch and a dirty defect classification branch.

[0015] In one or more embodiments of the application, a defect-process database is further included, and the application layer is configured to automatically match and conduct correlation analysis on the defect classification results and the production process parameters stored in the database to realize quality tracing.

[0016] Beneficial effects: Compared with the prior art, the EL cell defect intelligent classification method based on multi-model collaborative reasoning in the embodiment of the application can process different types of defects through a three-level collaborative reasoning architecture, accurately locate key defects such as hidden cracks through the parallel design of YOLOv8 and the DDB-Inception classification model, and finely distinguish confused defects such as black edges and dirt, and the comprehensive classification accuracy is more than 98%. At the same time, combined with the mixed precision reasoning and the dynamic load balancing algorithm, the processing throughput can reach 1200 pieces per minute, which greatly improves the detection efficiency.

[0017] The application integrates a defect-process database, can automatically associate the detected defect types with the production process parameters upstream, provides data support for process engineers to quickly locate the root cause of the problem, realizes a closed loop from discovering defects to solving problems, and effectively reduces the rate of defective products.

[0018] The DDB-Inception classification model in the application can adaptively process input images of different features through a dynamic routing mechanism and differential processing of double branches, effectively solves the sample imbalance problem, and improves the recognition ability of minority class defects such as dirt. The whole scheme does not need to frequently perform manual parameter adjustment like traditional methods, and has good generalization ability for battery pieces of different processes.

[0019] The application adopts a modular networking scheme of “distributed collection + centralized processing”, which can flexibly adapt to the scene of multiple production lines in parallel. The system design considers special production conditions such as line borrowing of a sorter, retains a manual selection mechanism, and reserves expansion space for the hardware architecture, and has good scalability and industrial applicability. BRIEF DESCRIPTION OF DRAWINGS

[0020] In order to make the technical solutions in the embodiments of the present application or the prior art clearer, the accompanying drawings needed in the embodiments or prior art description will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0021] Figure 1 is a flow chart of the EL cell defect intelligent classification method in an embodiment of the present application; Figure 2 is a technical architecture schematic diagram in an embodiment of the present application; Figure 3 is a networking scheme schematic diagram in an embodiment of the present application; Figure 4 is a dynamic reasoning flow chart in an embodiment of the present application; Figure 5 is a DDB-Inception classification model core module schematic diagram in an embodiment of the present application; Figure 6 is a software system main interface schematic diagram in an embodiment of the present application; Figure 7 is a software system parameter configuration interface schematic diagram in an embodiment of the present application; Figure 8 is a software system summary interface schematic diagram in an embodiment of the present application. DETAILED DESCRIPTION

[0022] In order to make those skilled in the art better understand the technical solutions in the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings of the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor should be within the scope of protection of the present application.

[0023] In view of the problems in the background art, as shown in Figure 1 The present application introduces an EL cell defect intelligent classification method based on multi-model collaborative reasoning, including the following steps: Step 1, obtaining image data of an EL cell to be detected; Step 2, the image data is processed by three-level collaborative inference, which includes: performing a first coarse-grained screening, performing a preliminary defect classification on the image data by a parallel double model architecture, obtaining a first classification result; based on the first classification result, a second positioning analysis is performed, when the first classification result is a preset specific category, a positioning model is triggered to perform spatial positioning on the defect; performing a third fine-grained discrimination, when the first classification result is a preset fine classification category, a dedicated sub-class discrimination model is called to obtain a second fine classification result; Step 3, outputting the first classification result and / or the second fine classification result at the application layer, and performing correlation analysis with a preset defect-process database according to the classification result.

[0024] Referring to Figure 2 The intelligent classification system proposed by the present application adopts a three-layer technical architecture, which includes The device layer is located at the bottom layer and is composed of 10 offline sorting machines as data acquisition terminals. Each sorting machine is responsible for EL imaging of the battery piece on the production line and uploading the collected original PNG format image data through the network.

[0025] The computing power model layer is deployed on the central server; this layer includes: The computing power facility unit includes high-performance CPU (which can be i712700), large-capacity memory (which can be 128G) and professional GPU (which can be NVIDIA A4000) to provide hardware support for model inference.

[0026] The model inference module internally deploys the three-level collaborative inference algorithm in the present application, including the YOLOv8 model and the DDB-Inception classification model, wherein DDB stands for Dynamic Dual Branch.

[0027] The data management module is responsible for storage (using 1T solid state and 4T mechanical hard disk combination), integration and data backflow of image data, and supports retention of original data and result data for 30 days for quality traceability.

[0028] The data analysis module includes data labeling, feature engineering and other functions, and is used for continuous optimization and iteration of the model.

[0029] The application layer is the user interaction layer, which provides a graphical user interface (GUI) developed based on PyQt5. This layer realizes defect classification result display, visual analysis, report output and other functions, and designs permission management to ensure that the first-level quality department is responsible for operation and the second-level process department can only view the results.

[0030] Referring to Figure 3As shown, the networking scheme of the system adopts the "distributed collection and centralized processing" mode. The 10 sorting machines are directly connected to the central server through the switch, forming a local area network. The EL image collected by the sorting machine is transmitted to the server through Ethernet, and the actual transmission speed can reach 50 MB / s, meeting the data volume requirement of processing more than 5000 pictures per hour. The current server is deployed with a single A4000 card, which can complete the processing of 15000 pictures (including first and second classification) in 11 minutes. The scheme specially designed a manual selection mechanism for non-standard production scenes such as "sorting machine borrowing lines", allowing operators to manually specify the data source on the software interface, enhancing the flexibility of the system.

[0031] In further embodiments, the three-level collaborative classification architecture and dynamic inference process in the present application, as shown in Figure 4 includes the following steps: The input is an EL cell image to be detected input_img.

[0032] First-level coarse-grained screening: This stage adopts a dual-model parallel architecture to simultaneously infer the pre-processed image tensor image_tensor and yolo_image.

[0033] YOLOv8 model: Specialized for detecting "EL hidden cracks" and "EL similar hidden cracks", two high-priority defects. The confidence threshold is set to 0.3. If any hidden crack or similar hidden crack frame with a confidence higher than 0.3 is detected, the system will prepare to return the result according to the preset priority (hidden cracks are prioritized over similar hidden cracks).

[0034] DDB-Inception classification model: Processes the remaining 11 defect categories (such as black edges, dirt, concentric circles, etc.). The model outputs the probability of each category and calculates the difference ΔP between the TOP1 and TOP2 probabilities. A dynamic threshold α (for example, α = 0.5) is used for judgment. Only when ΔP > α is the prediction result considered reliable.

[0035] Priority interruption mechanism: In the inference process, defect categories are assigned priorities: concentric circles > hidden cracks > similar hidden cracks > others. When a defect with high priority is detected, lower-level defects are no longer detected. If "EL concentric circles" is detected in the DDB-Inception classification model, the inference process will be immediately interrupted, and "EL concentric circles" will be returned as the final result. If "EL hidden cracks" with a confidence > 0.3 is detected in the YOLOv8 model, the inference process will be immediately interrupted, and "EL hidden cracks" will be returned. If no hidden cracks are detected, but "EL similar hidden cracks" are detected, the result will be temporarily stored, and the inference of the DDB-Inception classification model will continue, and finally the results will be comprehensively judged.

[0036] Second-level positioning analysis: This stage is implemented by the YOLOv8 model. When the first-level classification result is "EL hidden crack" or "EL class hidden crack", the detection box coordinates output by YOLOv8 are the position information of the defect, which can be used for subsequent size measurement or repair guidance.

[0037] Third-level fine-grained discrimination: This stage is a conditional trigger design, which is activated only when the first-level classification result is "EL black edge_black corner" or "EL dirty", to avoid unnecessary calculations.

[0038] If the first-level result is "EL black edge_black corner", the system will reuse the generated image_tensor and call the dedicated black edge subdivision model (model2_black) to further distinguish it into boat tooth black edge, black edge_black corner up and down, and black edge_black corner messy, etc.

[0039] If the first-level result is "EL dirty", the dirty subdivision model (model2_dirty) is called to subdivide it into dirty film, dirty leakage, dirty backflow, oxidized dirty, and dirty messy, etc.

[0040] These special sub-models have more targeted training data, and their accuracy in sub-classification can reach more than 85%. In this embodiment, the black edge subdivision model and the dirty subdivision model can be Inception_v3 classification models.

[0041] The entire process is optimized through a dynamic load balancing algorithm, for example, the models of different branches (model2_black and model2_dirty) are loaded independently, GPU resources are allocated on demand, and about 20% of the repeated calculation overhead is reduced by reusing image tensors.

[0042] The executable code corresponding to the three-level cooperative classification algorithm in this embodiment is as follows: #Input: input_img (image file path or PILImage object) #Output: (first-level classification result, second-level fine classification result) FUNCTION imageInfer(input_img): #1. Image preprocessing image_tensor, yolo_image <- preprocess_image(input_img) # Standardization + tensor conversion [1]() #2. First-level model inference (Inception_v3) WITH torch.no_grad(): outputs ← model1.predict(image_tensor, type=1) # output 11 class probabilities [1]() one_stage_result ← class_list1[outputs[0]] # take the highest confidence class # 3. Priority interrupt judgment IF one_stage_result == "EL concentric circle": RETURN(one_stage_result, NextInfer(one_stage_result)) # skip subsequent inference [1]() # 4. YOLOv8 hidden crack detection (parallel verification) yolo_results ← modelyolo.predict(yolo_image, conf=0.3) IF yolo_results exist detection box: FOR each detection box: cls_name ← class_list0[box class] conf ← box confidence # Hidden crack priority logic IF cls_name == "EL hidden crack" AND conf > 0.3: RETURN(cls_name, "") # immediately return hidden crack result [1]() ELSEIF cls_name == "EL class hidden crack" AND conf > 0.3: temporarily stored as a candidate result IF there is a valid hidden crack / class hidden crack result: RETURN(final hidden crack class, "") # 5. Default return first-level result + trigger second-level classification RETURN(one_stage_result, NextInfer(one_stage_result)) FUNCTION NextInfer(pred_class): # 6. Second-level fine-grained classification IF pred_class == "EL black edge_black corner": WITH torch.no_grad(): outputs←model2_black.predict(image_tensor, type=2)#Black-bordered subclass model[1]() RETURNclass_list2_black[outputs[0]]#Black border for stern teeth / top and bottom, etc. ELSEIFpred_class=="ELdirty": WITHtorch.no_grad(): outputs←model2_dirty.predict(image_tensor, type=3)#dirty subclass model[1]() RETURNclass_list2_dirty[outputs[0]]# Backflow / oxidation of dirt, etc. ELSE: RETURN #No need for further subdivision In a further embodiment, such as Figure 5 As shown, the DDB-Inception classification model includes a dynamic routing mechanism and a dual-branch structure. The backbone network (based on Inception_v3) is responsible for extracting multi-scale features of the image, generating a 2048-dimensional feature vector x. The routing controller receives this feature vector x and generates a class weight vector through two fully connected layers and an activation function. This process introduces a learnable parameter normal_boost (initial value 2.5) to dynamically amplify the weights of regular classes to address the problem of insufficient learning of minority class samples such as dirt.

[0043] The adjustment formula for the weight of the regular defect category is as follows: ； ； in, x The 2048-dimensional feature vector output by the backbone network represents the multi-scale defect features of the EL image of the battery cell. , This is the weight matrix, used for linear transformations; ∈ : Compress high-dimensional features to 512 dimensions to extract key defect features; ∈ Generate category weight vectors corresponding to different defect types; b 1. b2 is a bias term to enhance model flexibility; Hardswish is an activation function balancing non-linear expressiveness and computational efficiency, used to handle complex defect patterns in EL images.

[0044] σ Sigmoid represents a function that normalizes weights to [0, 1], indicating the priority of each defect class. normal boost is an initial learnable parameter used to dynamically boost the weight of regular defects.

[0045] In EL detection, the number of dirty samples is small but highly variable (e.g., different shapes, colors), while regular defects (e.g., scratches) have relatively stable features and require high-precision detection. By amplifying the weight of regular classes through normal_boost, we can avoid overfitting to dirty noise by the model.

[0046] The dynamic routing mechanism adaptively allocates weights based on input image features, such as preferentially allocating high-intensity areas to crack branches.

[0047] In further embodiments, the dual-branch structure includes: The regular branch is responsible for handling all defect classes except dirty. The Softmax layer of this branch uses a very low temperature parameter τ (e.g., 0.05), resulting in a very sharp probability distribution that helps the model make decisions with high confidence and reduces confusion among regular defects.

[0048] The dirty branch is designed specifically for identifying various dirty samples. This branch uses a higher temperature parameter τ (e.g., 0.8) to make the probability distribution smoother, improving the model's tolerance for dirty samples with diverse shapes and ambiguous features. Additionally, during model backpropagation training, the gradient of this branch is attenuated, which acts as a form of regularization to prevent overfitting due to the complex and variable nature of dirty samples, while also preventing excessive influence of dirty features on shared network layers.

[0049] Dynamic weight calculation and feature allocation includes: Weight generation: The router outputs weights with class_num dimensions, with the first class_num-1 dimensions corresponding to regular classes and the last dimension corresponding to the dirty class Weight enhancement: Regular class weights are multiplied by the learnable coefficient normal_boost Temperature adjustment principle: Adjust the temperature parameter τ Control the sharpness of the Softmax probability distribution of logits: ; where logits is the model's prediction for the i-th class of the j-th sample, and τ is the temperature parameter. iThe original output score of the class defect, used to reflect the confidence of the class defect in the EL image; τ For the temperature parameter, in the conventional branch τ = 0.05; In the dirt branch τ = 0.8, a higher temperature smoothing probability distribution to tolerate low confidence dirt samples to improve fault tolerance; P(y = i) The final classification probability, whose sharpness directly affects the strictness of EL detection.

[0050] Where: low τ Only strong light emitting defects are identified, suitable for key defect judgment; high τ Weak light emitting dirt may also be included, suitable for complex background noise.

[0051] Conventional branch: low τ Enhance the distinction between classes to avoid confusion between cracks and normal areas.

[0052] Dirt branch: high τ Mitigate false positives due to dirt diversity, and feature interference suppression in conjunction with gradient attenuation technology.

[0053] Through this design, the model can intelligently decide which branch to allocate more computing resources and attention according to the content of the input EL image, thereby significantly improving the classification performance on complex and unbalanced defect data while maintaining high computational efficiency.

[0054] In further embodiments, with reference to Figure 6 , Figure 7 and Figure 8 , the application provides a complete EL cell piece sub-image and aggregation software.

[0055] As shown in Figure 6 , the main interface can be used for users to select the sorter number, date, shift and time period, and click the button to pull the original image from the server. The system then automatically executes the above classification reasoning process and displays the results in real time in the interface table.

[0056] As shown in Figure 7 , it is a parameter configuration interface, which is used to configure the network parameters such as the names, IP addresses, login accounts of 10 sorters, to ensure that the system can correctly connect to each data acquisition terminal.

[0057] As shown in Figure 8As shown, it is a summary interface, the user can check one or more batches of test results (by line, time, etc.), the system will automatically generate statistical report, calculate the frequency of occurrence of various defects, such as table 1 and table 2 shown in the first and second classification result summary table. These report data can be directly used for quality analysis and process improvement. For example, when the report shows that the "boat tooth black edge" defect rate of a certain production line is abnormally high, the process department can quickly associate the corresponding graphite boat state and conduct investigation, realizing data-driven quality management.

[0058] Table 1. EL defect subgraph of battery piece and summary system first-level subgraph result table

[0059] Table 2. EL defect subgraph of battery piece and summary system second-level subgraph result table

[0060] In summary, the application provides an efficient, accurate and practical EL battery piece defect intelligent classification solution through the innovative multi-model collaborative reasoning architecture, the optimized dynamic reasoning process and the deep integration with the production process, effectively solves many pain points in the prior art, and has extremely high industrial application value.

[0061] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the application, and not to limit them; although the application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the application.

[0062] Those skilled in the art should understand that the embodiments of the application can be provided as a method, a system or a computer program product. Therefore, the application can adopt a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware aspects. Moreover, the application can adopt the form of a computer program product implemented on one or more computer usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer usable program code.

[0063] The present application is described in reference to the flowchart and / or block diagrams of the methods, apparatus (systems) and computer program products according to embodiments of the application. It will be understood that each block of the flowchart and / or block diagrams, and combinations of blocks in the flowchart and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks.

[0064] These computer program instructions can also be stored in a computer- readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the function specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks.

[0065] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks.

[0066] It will be apparent to those skilled in the art that the application is not limited to the details of the above-exemplified embodiments and that the present application can be implemented in other specific forms without departing from the spirit or essential character thereof. The embodiments are therefore to be considered in all respects as illustrative and not restrictive, the scope of the application being indicated by the appended claims rather than by the foregoing description, and all changes which come within the meaning and range of equivalency of the claims are therefore intended to be embraced therein. No limitation is intended to the scope of the claims based on any embodiment illustrated in the specification.

[0067] Furthermore, it should be appreciated that a single independent technical solution is not contained in each embodiment, and the description of the specification is only for the sake of clarity, and the skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can be combined appropriately to form other embodiments that can be understood by the skilled in the art.

Claims

1. A multi-model collaborative inference-based EL cell defect intelligent classification method, characterized in that, The method comprises the following steps: obtaining image data of an EL cell to be detected; performing three-level collaborative inference processing on the image data, which comprises: performing first-level coarse-grained screening, performing preliminary defect classification on the image data through a parallel double-model architecture to obtain a first-level classification result; performing second-level positioning analysis based on the first-level classification result, triggering a positioning model to perform spatial positioning on the defect when the first-level classification result is a preset specific category; and performing third-level fine-grained discrimination, calling a dedicated sub-category discrimination model to obtain a second-level fine-grained result when the first-level classification result is a preset fine-grained category; outputting the first-level classification result and / or the second-level fine-grained result and performing correlation analysis on the classification result and a preset defect-process database.

2. The multi-model collaborative inference-based EL cell piece defect intelligent classification method according to claim 1, characterized in that, The parallel double-model architecture comprises: a first model for detecting high-priority hidden cracks or hidden crack-like defects and outputting the position coordinates thereof; a second model for classifying the remaining preset defect categories; wherein, when the first model detects a hidden crack defect, the result is determined preferentially and the subsequent inference process can be interrupted. 3.The EL cell piece defect intelligent classification method based on multi-model collaborative reasoning according to claim 1, characterized in that, The preset fine-grained category comprises "EL black edge_black corner" and "EL dirt", wherein: when the first-level classification result is "EL black edge_black corner", a black edge fine-grained model is called to discriminate it into one of boat-tooth black edge, black edge_black corner up and down, and black edge_black corner disorder; when the first-level classification result is "EL dirt", a dirt fine-grained model is called to discriminate it into one of dark film, plating leakage, reverse plating, oxidation dirt, and dirt disorder.

4. The multi-model collaborative inference-based EL cell piece defect intelligent classification method according to claim 2, characterized in that, The second model comprises an Inception_v3 classification model and a router connected to the Inception_v3 classification model, wherein the router integrates a dynamic routing mechanism and a double-branch structure; The dynamic routing mechanism generates a category weight vector by a routing controller according to the image features extracted by the backbone network, and dynamically allocates the features to different processing branches. The double-branch structure comprises a regular defect classification branch and a dirt defect classification branch.

5. The multi-model collaborative reasoning-based EL cell piece defect intelligent classification method according to claim 4, characterized in that, The dynamic routing mechanism further comprises a learnable weight adjustment parameter normal_boost for dynamically adjusting the weight proportion of the regular defect category; the gradient of the dirt defect classification branch is subjected to attenuation processing in the model training process to suppress feature interference and prevent overfitting.

6. The multi-model collaborative reasoning-based EL cell piece defect intelligent classification method according to claim 5, characterized in that, The adjustment formula of the weight of the regular defect category is: ； ； wherein, x a 2048-dimensional feature vector output by the backbone network, representing multi-scale defect features of the battery piece EL image; , is a weight matrix used for linear transformation; ∈ For compressing high-dimensional features to 512 dimensions, extracting key defect features; ∈ for generating a category weight vector corresponding to different defect types; b 1、 b 2 is a bias term, enhancing model flexibility; Hardswish is an activation function that balances the non-linear expression ability and the calculation efficiency, and is used to process complex defect patterns in EL images. σ represents a sigmoid function that normalizes the weights to [0, 1] representing the priority of each class of defects; normal_boost are initial learnable parameters used to dynamically boost the weights of regular defects.

7. The multi-model collaborative reasoning-based EL cell piece defect intelligent classification method according to claim 4, characterized in that, The regular defect classification branch and the dirt defect classification branch use different temperature parameters τ to adjust the logits output by the model to control the sharpness of the Softmax probability distribution, wherein the temperature parameter of the regular defect classification branch is lower than that of the dirt defect classification branch; wherein, the formula for controlling the sharpness of the Softmax probability distribution of logits by the temperature parameter τ comprises: ； where logits is the original output score of the model for the i-th class of defects, which is used to reflect the confidence of the i-th class of defects in the EL image. i the original output score of the model for the i-th class of defects, which is used to reflect the confidence of the i-th class of defects in the EL image. τ = 0.05 for the temperature parameter, where the regular branch is τ = 0.05; In the dirty branch τ = 0.8, higher temperature smooth probability distribution to tolerate low confidence dirty samples to increase fault tolerance; P(y=i) The sharpness of the final classification probability directly influences the stringency of the EL detection.

8. A multi-model collaborative inference-based EL cell defect intelligent classification system, characterized in that, comprises: An equipment layer comprising at least one data acquisition terminal for collecting image data of an EL cell to be detected; A computing power model layer connected with the equipment layer, configured to receive and process the image data, and comprising the following inside: A first coarse-grained screening module configured to preliminarily classify defects in the image data through a parallel double-model architecture to obtain a first classification result; A second positioning analysis module configured to perform spatial positioning of defects when the first classification result is a preset specific category; A third fine-grained discrimination module configured to call a special sub-class discrimination model when the first classification result is a preset sub-category to obtain a second sub-classification result; An application layer connected with the computing power model layer, configured to output the first classification result and / or the second sub-classification result and provide a human-computer interaction interface.

9. The system of claim 8, wherein, The double-model architecture in the first coarse-grained screening module comprises: A first model unit that can adopt a YOLOv8 target detection model to detect high-priority hidden cracks or crack-like defects; A second model unit that can adopt a DDB-Inception classification model to classify the remaining preset defect categories, the DDB-Inception classification model comprising an Inception_v3 backbone network, a dynamic routing controller, and a regular defect classification branch and a contamination defect classification branch.

10. The system of claim 8, wherein, Further comprising a defect-process database, and the application layer is configured to automatically match and conduct correlation analysis on the defect classification result and the production process parameters stored in the database to realize quality tracing.

Images