Automatic screener of images of liquid-cooled tube connectors, automatic screening method

Abstract

The application provides an automatic screening device and method for images of a liquid cooling pipe connector, and relates to the field of battery management, and comprises an image acquisition module configured to acquire image data of the liquid cooling pipe connector; a processing module configured to intercept the image data to retain a connector image containing the liquid cooling pipe connector in the image data; determine a connector type according to the connector image; call a corresponding quality inspection prompt word according to the connector type; input the quality inspection prompt word and the connector image into a preset multimodal large model to obtain a first quality inspection result output by the multimodal large model, wherein the quality inspection prompt word comprises a quality inspection index corresponding to the connector type and does not comprise a guide word of the quality inspection result; and adjust the quality inspection prompt word according to the first quality inspection result to improve the quality inspection accuracy. The optimization and iteration of the prompt word are realized, and the accuracy of visual quality inspection of a lithium battery liquid cooling pipe quick plug connector is improved.

Description

Technical Field

[0001] This invention relates to the field of battery management, and more specifically, to an automatic image screening device and method for liquid cooling pipe connectors. Background Technology

[0002] With the rapid development of new energy vehicles and the energy storage industry, the thermal management safety of lithium batteries has become a core focus of industry attention. As a key component of battery pack thermal management, the liquid cooling system's sealing and connection reliability directly determine the operational safety of the battery system. The quick-connect fitting for the liquid cooling pipe is a crucial connecting component in the liquid cooling circuit, and its insertion quality directly affects the sealing performance of the cooling circuit. Poor fitting can lead to coolant leakage or insufficient flow, potentially causing localized thermal runaway and, in severe cases, even battery pack fire.

[0003] In related technologies, during the production process of lithium batteries, the quick-connect fittings of liquid cooling pipes are typically subjected to full-volume, high-precision visual quality inspection.

[0004] However, visual quality inspection methods in related technologies suffer from problems such as inaccurate recognition and low efficiency. Summary of the Invention

[0005] This application provides an automatic screening device and method for images of liquid cooling pipe connectors, in order to at least solve the problems of inaccurate identification and low efficiency in visual quality inspection of quick-connect lithium battery liquid cooling pipe connectors in related technologies.

[0006] One aspect of this application provides an automatic screening device for images of liquid cooling pipe connectors, comprising:

[0007] The image acquisition module is used to acquire image data of the liquid cooling pipe connector;

[0008] The processing module is used to crop the image data to retain the connector image containing the liquid cooling pipe connector; determine the connector type based on the connector image; call the corresponding quality inspection prompt word based on the connector type; input the quality inspection prompt word and the connector image into a preset multimodal large model to obtain a first quality inspection result output by the multimodal large model, wherein the quality inspection prompt word includes quality inspection indicators corresponding to the connector type but does not include guiding words for the quality inspection result; adjust the quality inspection prompt word based on the first quality inspection result to obtain an adjusted quality inspection prompt word, which is used to input the preset multimodal large model to perform quality inspection recognition on the connector image.

[0009] According to another embodiment of this application, a battery management system is also provided, including an automatic image screening device for liquid-cooled pipe connectors. The automatic image screening device for liquid-cooled pipe connectors includes: an image acquisition module for acquiring image data of liquid-cooled pipe connectors; a processing module for cropping the image data to retain connector images containing liquid-cooled pipe connectors; determining the connector type based on the connector image; using a multimodal large model to perform quality inspection identification on the connector image based on the connector type and calling corresponding quality inspection prompts to obtain a first quality inspection result, wherein the quality inspection prompts include quality inspection indicators corresponding to the connector type but do not include guiding words for quality inspection results; and adjusting the quality inspection prompts based on the first quality inspection result to improve the quality inspection accuracy.

[0010] According to another embodiment of the present application, a battery pack is also provided, including the battery management system of the above embodiment.

[0011] According to yet another embodiment of the present application, an electric vehicle is also provided, including the battery pack of the above embodiment.

[0012] In this embodiment, images of liquid cooling pipe connectors in diverse scenarios are acquired through an image acquisition module. Then, a processing module crops the image data, improving the effective resolution of key features and removing background noise interference. Next, corresponding quality inspection prompts are invoked based on the identified connector type, achieving differentiated and accurate quality inspection for different connector types and avoiding feature confusion caused by generic prompts. A multimodal large model is used to perform quality inspection identification on the connector images based on the corresponding quality inspection prompts for the connector type, obtaining a first quality inspection result. The quality inspection prompts include quality inspection indicators corresponding to the connector type but do not include guiding words for the quality inspection result. Through type-specific quality inspection prompts and single-round inference without injected quality inspection results, the accuracy and anti-interference capability of the model's judgment are improved. Finally, the quality inspection prompts are adjusted based on the first quality inspection result, thereby achieving optimization and iteration of the prompts and improving the accuracy of visual quality inspection of lithium battery liquid cooling pipe quick-connect connectors. Attached Figure Description

[0013] Figure 1 This is one of the flowcharts for an automatic image screening device for a liquid cooling pipe connector according to an embodiment of this application;

[0014] Figure 2 This is a second flowchart of an automatic image screening device for a liquid cooling pipe connector according to an embodiment of this application;

[0015] Figure 3 This is the third flowchart of an automatic image screening device for a liquid cooling pipe connector according to an embodiment of this application;

[0016] Figure 4 This is the fourth flowchart of an automatic image screening device for a liquid cooling pipe connector according to an embodiment of this application;

[0017] Figure 5 This is the fifth flowchart of an automatic image screening device for a liquid cooling pipe connector according to an embodiment of this application;

[0018] Figure 6 This is flowchart six of an automatic image screening device for a liquid cooling pipe connector according to an embodiment of this application;

[0019] Figure 7 This is flowchart seven of an automatic image screening device for a liquid cooling pipe connector according to an embodiment of this application;

[0020] Figure 8 This is one of the complete screening flowcharts of an automatic image screening device for a liquid-cooled pipe connector according to an embodiment of this application;

[0021] Figure 9 This is a block diagram of an automatic image screening device for a liquid cooling pipe connector according to an embodiment of this application;

[0022] Figure 10 This is the second complete flowchart of an automatic screening device for images of a liquid-cooled pipe connector according to an embodiment of this application. Detailed Implementation

[0023] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0024] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0025] The following appropriately discloses an embodiment of a battery according to this application. However, unnecessary detailed descriptions may be omitted. For example, detailed descriptions of well-known matters and repetitive descriptions of practically identical structures may be omitted. This is to avoid making the following description unnecessarily lengthy and to facilitate understanding by those skilled in the art. Furthermore, the following description is provided to enable those skilled in the art to fully understand this application and is not intended to limit the subject matter of the claims.

[0026] The battery in this application is a secondary battery, also known as a rechargeable battery or storage battery, which refers to a battery that can be used again after being discharged by recharging to activate the active materials.

[0027] Typically, a secondary battery includes an electrode assembly, an electrolyte, and an outer casing. The electrode assembly consists of a positive electrode, a negative electrode, and a separator. The electrode assembly and electrolyte are assembled inside the outer casing. During charging and discharging, active ions (such as lithium ions) move back and forth between the positive and negative electrodes, inserting and extracting. The separator, positioned between the positive and negative electrodes, primarily prevents short circuits while allowing active ions to pass through. The electrolyte, located between the positive and negative electrodes, mainly serves to conduct active ions.

[0028] This embodiment provides an automatic image screening device for liquid cooling pipe connectors, including: an image acquisition module and a processing module, wherein:

[0029] The image acquisition module is used to acquire image data of the liquid cooling pipe connector.

[0030] The image acquisition module can include an industrial camera or a handheld mobile camera to acquire visual information about the quick-connect liquid cooling pipe connectors on the lithium battery production line or at the repair station. The liquid cooling pipe connector is a critical sealing component in the battery thermal management system, and its connection status directly affects the reliability of the cooling circuit. The image data includes the connector body, the port short tube, the locking structure, and the surrounding environmental background.

[0031] Specifically, data acquisition methods can be divided into two modes: fixed-station acquisition and handheld mobile acquisition. Fixed-station acquisition utilizes high-definition industrial cameras installed on the production line, automatically triggering shooting as the conveyor belt passes by; handheld mobile acquisition allows quality inspectors to use smartphones or dedicated handheld cameras to photograph connectors in confined spaces from different angles and distances. The acquired raw image data typically contains rich pixel information, but also includes a significant amount of background noise unrelated to quality inspection, such as adjacent pipes, labels, cables, and environmental reflections.

[0032] Liquid cooling pipe connectors are quick-connect components used to connect battery pack liquid cooling piping systems. Common types include: Round Push-Fit type: A cylindrical black plastic elbow housing that pushes onto a silver metal or transparent plastic cooling port short pipe. Normal characteristics include a red locking ring at the base (invisible after insertion), colored alignment marks on the housing surface, and some models have a wide, permanent colored ring. Failure modes include incomplete insertion of the housing, protruding red locking ring, and exposed grooved ring. Rectangular Clip-On type: An angular / rectangular black housing that snaps around a silver cylindrical port short pipe from the side. Normal characteristics include a red snap on the side of the connection (invisible after insertion) and blue / green / yellow stripes on the housing surface. Failure modes include incomplete insertion of the housing, silver short pipes visible on both sides, and a separate red element on the side of the connection. The vertical push-up clip features a cylindrical black housing that snaps onto a short silver port tube extending from the top. The housing has a rectangular window slot. Normally, the red latch retracts into the housing after being in place. Failure modes include a large gap between the housing and the port, and the red latch extending beyond the diameter of the housing.

[0033] like Figure 1 As shown, the processing module is used to perform the following steps:

[0034] Step S100: Capture the image data to retain the connector image containing the liquid cooling pipe connector.

[0035] The processing module receives the raw image data and then performs region-of-interest (ROI) localization and cropping operations. It removes background interference and extracts only the effective image region containing the key connector structure, thus solving the problem of diluted key feature resolution caused by full-image input. In the original image, the key detection area of ​​the connector only occupies 5%-15% of the image, and severe background noise interferes with the model's judgment. The cropped connector image magnifies the key features, increasing the proportion of the key area to over 70% of the image, improving the effective resolution by approximately 4 to 6 times. This makes subtle mating differences (such as tiny gaps and protruding latches) much clearer.

[0036] Step S110: Determine the connector type based on the connector image.

[0037] The processing module analyzes the structural features of the connector image and identifies its specific mechanical structure type in order to load the matching quality inspection logic.

[0038] Specifically, the processing module utilizes the visual recognition capabilities of a multimodal large language model to perform morphological analysis on connector images. The model classifies connectors based on their geometry, latching direction, and port connection method. For example, if an image shows a cylindrical black shell pushed into a silver port, it is classified as a round push-fit; if a rectangular shell is shown to be latched from the side, it is classified as a rectangular clip-on; and if a cylindrical shell is shown to be latched vertically from below, it is classified as a push-up clip. Different types of connectors exhibit significant differences in the visual characteristics of their normal state and failure modes. By identifying the connector type, differentiated quality inspection for specific structural types is achieved. Connectors with different structures have different failure modes (such as different red latch positions and different gap judgment standards). By pre-determining the type, the system can accurately match dedicated quality inspection prompts for subsequent steps, improving the applicability and accuracy of the quality inspection rules.

[0039] Step S120: Call the corresponding quality inspection prompt word according to the connector type.

[0040] Step S130: Input the quality inspection prompt and the connector image into the preset multimodal large model to obtain the first quality inspection result output by the multimodal large model.

[0041] The quality inspection prompts include quality inspection indicators corresponding to the connector type, but do not include guiding words for quality inspection results.

[0042] The processing module loads quality inspection prompts that match the identified connector type, inputs the connector image and prompts together into a multimodal large language model, and the model independently infers based on visual features to output structured quality inspection conclusions.

[0043] Specifically, the processing module determines the connector type and retrieves the corresponding quality inspection prompts from the knowledge base. For example, for elbow push-in connectors, the prompts emphasize the characteristics of the red locking ring on the base and the metal pipe groove; for rectangular side-clamp connectors, the prompts emphasize the visibility of the red side clip and the silver short pipe. The prompts clearly define the primary indicator (connection status OK / NG) and secondary indicators (red locking status, groove visibility), and include anti-confusion rules (such as distinguishing between design color strips and faulty clips). The prompts do not contain any known quality inspection conclusions, nor do they include any guiding text regarding the quality inspection conclusions. After receiving the image and prompts, the multimodal large model performs a single round of inference and outputs the first quality inspection result. Quality inspection indicators include primary and secondary indicators. The quality inspection prompts are used to instruct the multimodal large model to output quality inspection results based on the primary indicator, determine the primary indicator based on the secondary indicator, and not participate in the judgment of the primary indicator if the secondary indicator is not identifiable.

[0044] Step S140: Adjust the quality inspection prompts based on the first quality inspection result to improve the accuracy of quality inspection.

[0045] Among these features, the system can compare the first quality inspection result with the manually labeled true value, analyze the cause of the error, automatically modify or generate a new version of the quality inspection prompt, and perform a full backtest to verify it, thereby iterating the quality inspection prompt and improving the quality inspection accuracy.

[0046] In this embodiment, images of liquid cooling pipe connectors in diverse scenarios are acquired through an image acquisition module. Then, a processing module crops the image data, improving the effective resolution of key features and removing background noise interference. Next, corresponding quality inspection prompts are invoked based on the identified connector type, achieving differentiated and precise quality inspection for different mechanical structures and avoiding feature confusion caused by generic prompts. Based on the connector type, a multimodal large model is used to perform quality inspection identification on the connector images, obtaining a first quality inspection result. The quality inspection prompts include quality inspection indicators corresponding to the connector type but do not include guiding words for the quality inspection result. Through type-specific quality inspection prompts and single-round inference without injected quality inspection results, the accuracy and anti-interference capability of the model's judgment are improved. Finally, the quality inspection prompts are adjusted based on the first quality inspection result, thereby achieving optimization and iteration of the prompts and improving the accuracy of visual quality inspection of lithium battery liquid cooling pipe quick-connect connectors.

[0047] In one embodiment, such as Figure 2 As shown, the processing module is used to perform the following steps:

[0048] Step S200: Determine whether the first quality inspection result is correct based on the first quality inspection result, the connector image, and the quality inspection indicators of the connector type corresponding to the connector image.

[0049] The processing module receives the first quality inspection result output by the multimodal large language model. This result may include structured primary judgment indicators (such as connection status OK / NG) and auxiliary judgment indicators (such as red latch status, groove visibility, and housing positioning status). The processing module compares the primary judgment indicators in the first quality inspection result with the pre-labeled true values, and verifies the correctness of the judgment by combining the specific quality inspection indicator logic for the connector type.

[0050] Step S210: If the first quality inspection result is determined to be incorrect, the cause of the error is determined based on the first quality inspection result and the manual annotation result.

[0051] In this process, after confirming that the first quality inspection result is incorrect, the root cause of the difference between the model output and the true value is analyzed, and the error is classified into a specific error cause type so that targeted optimization strategies can be adopted.

[0052] For example, the error classification logic can be executed by combining the visual description and auxiliary indicator status in the first quality inspection result with the manual annotation results. For instance: if the model misclassifies a design feature (such as colored paint stripe) as a failure feature (such as a red buckle), it is classified as visual confusion. If the model forcibly infers an NG state when the feature is not visible, and the manual annotation is uncertain, it is classified as over-inference caused by angle occlusion; if the model wavers in boundary conditions (such as whether the buckle protrudes), it is classified as boundary ambiguity; if the model fails to identify the connector type, resulting in rule mismatch, it is classified as type unidentified; if the judgment logic itself has a flaw (such as incorrectly using auxiliary indicators as a rejection condition), it is classified as a logical flaw.

[0053] Step S220: Adjust the quality inspection prompt words according to the error reason.

[0054] Based on the identified error causes, the current quality inspection prompts are modified or new versions are generated to correct cognitive biases or logical flaws in the model. By mapping error causes to specific text modification strategies, the system can automatically repair model defects in specific scenarios. This attribution-based targeted adjustment ensures that each iteration effectively addresses the specific problems of the prompts. The adjusted prompts can more accurately capture key visual features and suppress interfering information, thereby gradually improving the model's generalization ability and judgment accuracy in complex situations.

[0055] In this embodiment, by comparing the first quality inspection result with the manual annotation result through the processing module, and combining the specific quality inspection indicators for the connector type, the correctness of the model's judgment can be accurately determined, solving the problems of low efficiency and difficulty in comprehensively evaluating model performance caused by manual re-inspection. Secondly, when an error is judged, the system does not blindly adjust, but analyzes the difference between the first quality inspection result and the manual annotation to determine the cause of the error, achieving error attribution from phenomenon to essence. Subsequently, based on the determined cause of the error, the system adopts targeted strategies to adjust the quality inspection prompt words, such as adding feature comparison descriptions to resolve confusion, or refining quantification standards to eliminate boundary ambiguity. This makes the optimization process of the prompt words have clear direction and logic, avoiding the randomness and inefficiency of manual debugging. Through continuous iterative optimization, more accurate quality inspection prompt words are obtained, enabling the multimodal large model to more accurately identify the insertion status of the liquid cooling pipe connector.

[0056] In one embodiment, such as Figure 3 As shown, the processing module is used to perform the following steps:

[0057] In step S300, if the first quality inspection result is "quality inspection passed" and the manual annotation result is "quality inspection failed", it is determined that the first quality inspection result is incorrect and the reason for the error is that the quality inspection indicator was missed.

[0058] The processing module compares the main judgment index in the first quality inspection result output by the multimodal large language model with the ground truth value of the manually labeled data. When the model judges it as qualified, but the manual label judges it as unqualified, the processing module defines this situation as a quality inspection index omission error. This is because the model failed to recognize the specific visual features that cause the unqualified data, such as failing to detect a small red buckle protrusion, failing to recognize the groove on the metal pipe, or failing to notice the small gap between the shell and the port. Therefore, the error type attribute is recorded as omission, so that prompt words can be used to strengthen these missed features in the future. By accurately identifying the problem of the model being too lenient or the feature sensitivity being insufficient, and by clearly defining such errors as omissions, the system can specifically enhance the model's visual sensitivity to failure features (such as red buckles and grooves).

[0059] Step S310: If the first quality inspection result is "quality inspection failed" and the manual annotation result is "quality inspection passed", determine that the first quality inspection result is incorrect and the reason for the error is misjudgment of the quality inspection indicator.

[0060] The processing module compares the first quality inspection result with the manually labeled true value. When the model determines it to be unqualified, but the manual label determines it to be qualified, the processing module defines this situation as a misjudgment of the quality inspection indicator. This indicates that the current model has misjudged, because the model mistakenly identifies normal design features (such as colored aligned paint stripes and permanent color rings) as failure features (such as red buckles), or because reflections or angular obstruction cause the model to over-infer, resulting in defects. Recording the error type attribute as misjudgment indicates that the model lacks the ability to distinguish interference features in the normal state, and the anti-confusion rules in the prompt words need to be optimized. The system accurately identifies the problem of the model being too sensitive or feature confusion. By clearly defining such errors as misjudgments, the system can add targeted rules to distinguish between design features and failure features in the prompt words. For example, it can clearly state that certain colored stripes are only design marks and not buckles, thereby reducing the false alarm rate.

[0061] In this embodiment, the processing module performs a detailed logical comparison between the first quality inspection result and the manually labeled result, establishing a clear error classification mechanism. This facilitates determining the optimization direction based on the cause of the error, adjusting the quality inspection prompts in a targeted manner, and improving the recognition accuracy.

[0062] In one embodiment, such as Figure 4 As shown, the processing module is used to perform the following steps:

[0063] Step S400: Determine the recognition results of each quality inspection indicator in the connector image output by the multimodal large model based on the current quality inspection prompt words.

[0064] Each quality inspection indicator refers to a detailed visual feature parameter used to comprehensively evaluate the connector mating quality, including the main indicator of connection status and auxiliary indicators such as the status of the red latch, the visibility of the groove, and the housing positioning status. The recognition result refers to the specific classification label or status description output by the multimodal large model for each quality inspection indicator after analyzing the image based on the current prompt word. By obtaining the independent recognition results of each quality inspection indicator, the system no longer only focuses on whether the final result is qualified, but can delve into the internal reasoning chain of the model to understand which specific visual feature the model deviated from, providing a data foundation for subsequent accurate error localization and prompt word optimization.

[0065] The processing module parses the structured output data generated by the multimodal large language model and extracts independent judgments for each specific visual feature in the connector image. The processing module receives data returned by the multimodal large language model, which includes not only the final connection status but also the recognition results of intermediate auxiliary indicators. The processing module extracts the specific field values ​​of each quality inspection indicator from this data. For example, for elbow push-in connectors, the processing module extracts two key auxiliary indicators: the red latch status and groove visibility, as well as the final main indicator of the connection status. These recognition results reflect the model's visual perception and understanding of specific physical features in the image under the guidance of the current prompt words.

[0066] Step S410: Compare the identification results of each quality inspection indicator with the manual annotation results of each quality inspection indicator to determine the target quality inspection indicator where the judgment error occurred.

[0067] The processing module compares the model's output of each quality inspection indicator with the ground truth values ​​annotated by human quality inspectors, identifying specific indicators with discrepancies and marking them as target quality inspection indicators. If all auxiliary indicators match the human annotations, and the main indicator also matches the human annotations, then there are no target quality inspection indicators, and the overall result is considered correct. Through item-by-item comparison, the system can precisely pinpoint which specific visual feature recognition led to the incorrect final quality inspection conclusion.

[0068] Step S420: Determine the defects in the current quality inspection prompts based on the target quality inspection indicators.

[0069] The prompt word defects include at least one of the following: visual confusion, over-judgment when features are not visible, blurred judgment boundaries, incorrect connector type identification, or judgment logic error. The processing module combines the error type of the target quality inspection indicator, the model's inference description, and the connector type to analyze the root cause of the indicator's identification error and classify it as a specific prompt word defect.

[0070] For example, if the target quality inspection indicator is a red latch state, and the error manifests as misjudging colored paint strips as red latches, then the prompt word defect is determined to be visual confusion. If the target quality inspection indicator is groove visibility or a red latch state, and the error manifests as the model forcibly inferring the presence or absence of a feature when it is invisible due to angular occlusion, which contradicts manual annotation, then the prompt word defect is determined to be overjudgment when the feature is invisible. If the target quality inspection indicator is a red latch state, and the error manifests as wavering between slightly protruding and completely flush, making it difficult to define, then the prompt word defect is determined to be ambiguous judgment boundary. If the error occurs when the model fails to correctly identify the connector type, resulting in the loading of incorrect dedicated rules, then the prompt word defect is determined to be connector type identification error. If all auxiliary indicators are correctly identified, but the judgment logic of the main indicator connection state violates the rules (e.g., judging it as unqualified based on only a vague auxiliary indicator), then the prompt word defect is determined to be judgment logic error.

[0071] Step S430: Adjust the quality inspection prompt words according to the defect of the prompt words.

[0072] Specifically, based on the identified defect type of the prompt words, the current quality inspection prompt words are modified in a targeted manner to generate a new version of the prompt words, thereby eliminating the defect and obtaining more accurate prompt words.

[0073] For example, if the defect is visually confusing, a comparative description of the design features and failure features can be added to the prompt. For instance, regarding the misjudgment of the red latch, it can be explained that the black housing above and below the wide colored strip is a design feature and is definitely not a locking indicator, and it can be emphasized that the red latch is a 3D raised plastic ring, not a flat painted strip, etc. If the defect is due to unclear judgment boundaries, the description can be refined and quantified in the prompt. For instance, specific visual reference standards such as "10mm+ exposed neck or significantly wider than the housing" can be added to reduce the space for subjective judgment.

[0074] In this embodiment, the processing module parses the identification results of each quality inspection indicator output by the multimodal large model and compares them item by item with the manually labeled true values. This allows for the precise identification of the target quality inspection indicator that leads to the error in the final quality inspection conclusion. Subsequently, the system combines the erroneous behavior of the target quality inspection indicator with the model's reasoning process to categorize the root causes of the errors into specific prompt word defects such as visual confusion, overjudgment when features are not visible, blurred judgment boundaries, incorrect connector type identification, or judgment logic errors. This establishes a precise mapping from indicator errors to prompt word defects. Finally, based on the determined prompt word defect types, the system executes targeted prompt word adjustment strategies, such as adding feature comparison descriptions, refining quantification standards, and adding uncertainty handling rules, thereby generating an optimized new version of prompt words and improving the quality inspection accuracy.

[0075] In one embodiment, such as Figure 5 As shown, the processing module is used to perform the following steps:

[0076] Step S500: Input the image data and preset recognition prompts into the multimodal large model to identify the key area containing the liquid cooling pipe connector in the image data.

[0077] The processing module utilizes a multimodal large language model, employing specially designed fixed template prompts to perform semantic understanding and spatial localization on the raw image data, thereby determining the specific location range of the liquid cooling pipe connector in the image. The processing module first acquires raw image data from the image acquisition module. Then, it constructs a fixed "region localization prompt," which contains only the localization task instruction and no quality control logic. The processing module inputs the raw image and the localization prompt into the multimodal large language model. Based on the prompt requirements, the model searches the image for visual features matching the description (such as a black plastic shell or metal port) and calculates the coordinates of a bounding box that tightly encloses the key region. The model outputs this coordinate sequence, which the processing module receives and parses to determine the spatial location of the key region in the raw image. This high-precision automatic localization of the key region is achieved by leveraging the semantic understanding capabilities of the multimodal large model.

[0078] Step S510: Extract the key region from the image data to obtain the connector image of the key region.

[0079] In this embodiment, the processing module inputs the original image data and preset fixed recognition prompts into a multimodal large model. Leveraging the semantic understanding capabilities of the large model, it accurately locates the key areas of the liquid cooling pipe connector, solving the problem of traditional visual algorithms struggling with location in complex backgrounds. Subsequently, the processing module crops the original image based on the location coordinates, generating a partial connector image containing only the connector body. The cropped connector image magnifies the key areas, effectively increasing the resolution by 4-6 times, and removing background noise, thus improving the accuracy of subsequent identification of minor mating defects.

[0080] In one embodiment, such as Figure 6 As shown, the processing module is used to perform the following steps:

[0081] Step S600: Receive image data of liquid cooling pipe connectors of various connector types as manually labeled sample data.

[0082] The processing module collects and stores image datasets covering various liquid cooling pipe connector structure types. These datasets form the basis of the sample library for subsequent multimodal large model prompt word optimization and validation.

[0083] Step S610: Receive the annotation information input by the user, annotate the sample data according to the annotation information, and obtain the manual annotation result of the sample data.

[0084] The processing module provides an interactive interface or data interface for quality inspectors to annotate sample images according to the work instructions, and stores the annotation results in a structured manner to form ground truth data for training and evaluation.

[0085] In this embodiment, various types of image data are first collected to construct a diverse and highly representative basic sample library, solving the problem of poor model generalization ability caused by insufficient samples from a single scene. Then, by receiving detailed annotation information input by quality inspectors according to the work instructions, the sample images are annotated to obtain the manual annotation results of the sample data, which facilitates subsequent comparison and verification.

[0086] In one embodiment, such as Figure 7 As shown, the processing module is used to perform the following steps:

[0087] Step S700: Input the test connector image into the multimodal large model.

[0088] Among them, the test connector image is the one whose quality inspection results have been confirmed to be correct.

[0089] The processing module selects specific test connector images from a validated sample library. These test connector images have been validated by previous steps, and their manual annotation results are consistent with the output results of the model under normal inference, ensuring the reliability of the test benchmark.

[0090] Step S710: Input counterfactual prompts into the multimodal large model to obtain the second quality inspection result.

[0091] Among them, counterfactual prompts include guiding words that indicate quality inspection results that contradict the confirmed quality inspection results of the test connector image.

[0092] The processing module constructs counterfactual cue words containing misleading information and inputs them along with the test connector image into the multimodal large language model. This induces the model to adopt incorrect textual conclusions against visual evidence, generating a second quality control result. By introducing adversarial testing, the text compliance defects of the multimodal large language model in visual quality control are quantified and exposed.

[0093] In step S720, if the second quality inspection result is inconsistent with the confirmed quality inspection result of the test connector image, adjust the quality inspection prompt words to increase the dependence of the multimodal large model on visual features.

[0094] The processing module compares the second quality inspection result with the confirmed quality inspection result. If the model is misled by counterfactual prompts and outputs incorrect conclusions, it determines that the current prompts are not strong enough to resist interference, and optimizes the prompts accordingly to strengthen the visual priority principle.

[0095] Specifically, the processing module first compares the second quality inspection result with the confirmed quality inspection result of the test connector image. If the second quality inspection result is inconsistent with the confirmed result, it indicates that the model has failed to resist the guidance of counterfactual prompts. At this point, the processing module determines that the current quality inspection prompts have a deficiency in low visual feature dependence. To correct this deficiency, the processing module can adjust the quality inspection prompts, adding instructions that emphasize the priority of visual evidence. After discovering the vulnerability of the model to text manipulation through counterfactual testing, the emphasis on the priority of visual features in the prompts is immediately enhanced. This optimization not only improves the accuracy of the model in normal inference and reduces the interference of text noise, but also enhances the security of the system during the operation and maintenance phase, preventing large-scale misjudgments caused by the accidental injection of incorrect labels.

[0096] In this embodiment, test connector images are first selected from verified correct samples to obtain benchmark test data. Then, an adversarial test scenario is constructed by inputting counterfactual cue words containing contradictory conclusions into the model to test whether the model will blindly follow the text when visual evidence conflicts with the textual cue. When a second quality inspection result is detected to be inconsistent with the confirmed correct result, it is determined that the current cue word lacks sufficient anti-interference capability, and the cue word is adjusted to increase the model's reliance on visual features and improve its ability to ignore erroneous textual guidance. This solves the problem that multimodal large language models are easily manipulated by textual cuees in visual quality inspection. The quality inspection cue words verified by counterfactual gating not only have high image recognition accuracy but also a certain degree of resistance to textual manipulation, thereby improving the stability and security of the automatic image screening device for liquid cooling pipe connectors.

[0097] In one embodiment, the processing module is used to input multiple different test connector images and counterfactual prompts into a multimodal large model to obtain multiple second quality inspection results output by the multimodal large model.

[0098] The test connector images refer to image data extracted from samples that have been verified as correct for security testing. Counterfactual cue words are textual guidance or erroneous conclusions that contradict the true nature of the images, used to test the model's robustness against interference. The second quality inspection result refers to the quality inspection judgment output by the multimodal large model based on textual guidance rather than purely visual features after receiving the counterfactual cue words. The processing module extracts a certain proportion of test images from the verified sample set, constructs counterfactual cue words containing erroneous conclusions for each image, and inputs both the image and the cue words into the multimodal large model, collecting the model's output results under misleading conditions.

[0099] Specifically, the processing module first randomly selects or chooses a representative subset of images from a historically validated sample library as a test set, ensuring that the test set covers different connector types, lighting conditions, and failure modes. For each test connector image in the test set, the processing module generates a corresponding counterfactual cue word, which explicitly contains a quality control conclusion that contradicts the image's true annotation (e.g., if the image's true value is "OK," the cue word guides the model to determine "NG" or provide an incorrect feature description). The processing module sends these images and corresponding counterfactual cue words in batches to the multimodal large model. Upon receiving the input, the model may ignore the true visual evidence and tend to follow the textual guidance, outputting a corresponding second quality control result.

[0100] The percentage of erroneous results is determined based on multiple secondary quality inspection results and confirmed quality inspection results from multiple different test connector images.

[0101] The processing module compares each second quality inspection result with the confirmed quality inspection result (true value) of the corresponding test connector image one by one, counts the number of inconsistent results, and calculates the proportion of this number to the total number of test samples.

[0102] Specifically, the processing module iterates through multiple stored second quality control results. For each result, the processing module looks up the confirmed quality control result (i.e., the manually labeled truth value, such as "OK" or "NG") for the corresponding test connector image. If the second quality control result is inconsistent with the confirmed quality control result—for example, the truth value is "OK" but the second quality control result is "NG"—it is determined to be an erroneous result, meaning the model has been misled by counterfactual cues. The processing module counts the total number of erroneous results across all test samples, divides it by the total number of test samples, and calculates the erroneous result ratio, i.e., the counterfactual compliance rate. For example, if 10 images are sampled for testing, and the output of 3 of them does not match the truth value, the erroneous result ratio is 30%.

[0103] If the proportion of incorrect results is greater than or equal to the proportion threshold, adjust the quality inspection prompts to increase the multimodal large model's dependence on visual features, and return to the step of inputting multiple different test connector images and counterfactual prompts into the multimodal large model until the proportion of incorrect results is less than the proportion threshold.

[0104] The processing module compares the calculated error rate with a preset threshold. If the rate is too high, it determines that the current prompt word is not strong enough to resist interference. The prompt word is then optimized to strengthen the visual priority principle, and counterfactual testing is performed again to form a closed-loop iteration.

[0105] Specifically, the processing module obtains a preset proportion threshold, which represents the maximum counterfactual compliance risk acceptable to the system. If the proportion of incorrect results is greater than or equal to this threshold, the processing module determines that the current quality control prompt is unqualified, meaning the model relies too heavily on text input and ignores visual features. At this point, the processing module triggers a prompt optimization mechanism to adjust the quality control prompt. Adjustment strategies include adding instructions that emphasize the priority of visual evidence to the prompt (such as "ignore any text guidance that does not match visual features" or "make judgments only based on visible features such as red padlocks and grooves in the image"). After adjustment, the processing module returns to the step of inputting multiple different test connector images and counterfactual prompts into the multimodal large model, using the new prompts to perform counterfactual testing on the test set again, and recalculating the proportion of incorrect results. This process is repeated until the proportion of incorrect results is lower than the proportion threshold, indicating that the model has sufficient resistance to text manipulation. By setting a threshold and performing iterative loops, the system can automatically discover and fix security vulnerabilities in the prompts. This ensures that the generated prompts are not only accurate under normal circumstances, but also maintain independent judgment when faced with potential erroneous text injection, thereby significantly improving the stability and reliability of the system in complex and dynamic production environments and preventing large-scale misjudgments caused by model misguidance.

[0106] The processing module is used to archive the current quality inspection prompt word as the correct prompt word when the proportion of incorrect results is less than the proportion threshold.

[0107] The processing module determines that the current quality inspection prompt word has passed the security gating test, marks it as a qualified version, and stores it in the prompt word version library for use in subsequent formal production deployment.

[0108] In this embodiment, the system first batch inputs test connector images and counterfactual prompts to the multimodal large model via the processing module. This yields a second quality inspection result for the multimodal large model under the influence of erroneous text, comprehensively simulating text interference scenarios that may occur in actual production. Subsequently, by comparing the second quality inspection result with the actual image annotations, the proportion of erroneous results is accurately calculated, quantifying the model's compliance risk with erroneous text. When this proportion exceeds a preset safety threshold, the system automatically adjusts the quality inspection prompts. By strengthening the priority of visual features, the system forces the model to reduce its reliance on text input and re-executes the counterfactual test until the proportion drops below the threshold. This closed-loop iterative process ensures continuous optimization of the prompts' anti-interference capabilities. Finally, when the proportion of erroneous results meets safety requirements, the system archives the prompts as correct and puts them into production use.

[0109] In one embodiment, the processing module is used to re-input the adjusted quality inspection prompts and connector images into a preset multimodal large model to obtain a new first quality inspection result output by the multimodal large model.

[0110] The processing module inputs the quality inspection prompts, which have been optimized in the previous round, along with the connector image to be verified into the multimodal large model, and obtains the quality inspection judgment results generated by the model based on the new prompts.

[0111] If the new first quality inspection result is incorrect, adjust the quality inspection prompt based on the new first quality inspection result, and return to the step of re-entering the adjusted quality inspection prompt and connector image into the preset multimodal large model until the new first quality inspection result is correct.

[0112] The processing module compares the new first quality inspection result with the manually labeled true value to determine whether the first quality inspection result is correct. If the first quality inspection result is incorrect, the quality inspection prompt words are adjusted in accordance with the method described in the above embodiment and re-entered into the multimodal large model for judgment until the new first quality inspection result is judged to be correct.

[0113] The processing module is used to archive the current adjusted quality inspection prompt as the correct prompt if the new first quality inspection result is correct.

[0114] In this embodiment, the optimized quality inspection prompts and connector images are first re-input into the multimodal large model via the processing module to obtain new quality inspection results, thereby verifying the effectiveness of the prompt adjustment. Subsequently, by comparing the new results with the human ground truth, if the first quality inspection result is incorrect, the prompts are adjusted, such as refining feature descriptions and strengthening visual priority logic, until the model outputs the correct first quality inspection result. The current adjusted quality inspection prompts are then archived as correct prompts.

[0115] In one embodiment, the processing module is used to record the number of iterations for the step of returning to the execution to re-enter the adjusted quality inspection prompts and connector images into the preset multimodal large model.

[0116] In the previous embodiment, when the processing module starts the closed-loop process for optimizing prompts for difficult cases, it initializes a counter and increments the counter after each execution of the action of re-inputting the adjusted quality inspection prompts and connector images into the preset multimodal large model, so as to record the number of optimization iterations that the current quality inspection prompts have undergone.

[0117] If the number of iterations reaches the preset iteration limit, stop using the multimodal large model to output the first quality inspection result.

[0118] The processing module compares the current number of iterations with the preset maximum allowed number of iterations, i.e., the iteration limit. If the counter value is greater than or equal to the limit, the current sample is determined to be an extremely difficult case, the automatic optimization process for the sample is terminated, and the output of the multimodal large model is abandoned as the final quality inspection conclusion.

[0119] The connector images are transmitted to the manual inspection station for quality inspection by humans.

[0120] The process involves sending the connector image to a manual inspection station, where senior quality inspectors visually assess the image and then feed the assessment results back to the system.

[0121] In this embodiment, the processing module first records the number of iterations for optimizing prompts for difficult examples, enabling quantitative monitoring of resource consumption during the optimization process. Then, the processing module compares the number of iterations with a preset iteration limit. Once the limit is reached, the current sample is determined to be an extremely difficult example that the system's automatic optimization cannot resolve, thus forcibly stopping the automatic output process of the multimodal large model, avoiding system stagnation and resource waste caused by individual complex samples. Finally, the processing module transmits these failed connector images to a manual review station for final visual quality inspection by senior quality inspectors, thereby avoiding multiple iterations for difficult-to-judge samples and improving the system's reliability and stability.

[0122] In one embodiment, the processing module is used to identify the geometry, snap-fit ​​orientation, and port connection method of the liquid cooling pipe connector based on the connector image.

[0123] The connector type is determined based on the geometry, snap-fit ​​direction, and port connection method of the liquid cooling pipe connector.

[0124] The processing module determines the connector type as elbow push-in if it determines that the connector includes a cylindrical elbow housing and the cylindrical elbow housing is pushed into the cooling port short pipe.

[0125] The processing module determines the connector type as rectangular side-clamp type if it determines that the connector includes an angular or rectangular housing and the angular or rectangular housing surrounds the cooling port short tube from the side with a snap fastener.

[0126] The processing module determines that the connector type is vertical top-clamp type if it determines that the connector includes a cylindrical housing and the cylindrical housing is snapped into the cooling port short pipe from bottom to top.

[0127] In this embodiment, the processing module first utilizes the visual capabilities of a multimodal large model to accurately extract fine-grained features such as geometric shape, snap-fit ​​direction, and port connection method from the connector image. Subsequently, based on these features, the processing module maps the specific visual information to one of three standard connector types—elbow push-in, rectangular side-clamp, and vertical top-clamp—using logical matching rules to determine the connector type.

[0128] In one embodiment, such as Figure 8As shown, the entire quality inspection process includes:

[0129] Step 1: Sample and task configuration.

[0130] Specifically, an labeled quality inspection sample set and task boundaries are constructed: n images of liquid-cooled connectors are collected, which can be acquired using a handheld workstation camera, covering different workstation scenarios such as the cold plate side, the cabinet side, and the cabinet connection point. Each image is labeled with different sample types, such as qualified, unqualified, and imaging limited. Each image is also manually annotated, including the true value of the main indicator of connection status, as well as auxiliary indicator annotations such as groove visibility and red latch status. The three connector types are configured with feature content: elbow push-in type, rectangular side-clamp type, and vertical top-clamp type.

[0131] Step 2: Locate the Region of Interest (ROI).

[0132] Specifically, a multimodal large model is independently invoked, along with fixed-template location prompts, to perform connector area identification and cropping on each original image. Only the key ROI area of ​​the black plastic shell and the port junction is retained, while irrelevant interference such as remote hardware, cables, and labels is removed. This can improve the focus and accuracy of subsequent quality inspection.

[0133] Step 3: Generate initial prompt words.

[0134] Specifically, the AI ​​agent automatically generates an initial version of the quality inspection prompts based on the background of the quality inspection task. The content includes complete information such as role settings, connector structure descriptions, main indicator judgment requirements, auxiliary indicator judgment rules, and output format constraints. This serves as the starting point for subsequent iterations and optimizations of the quality inspection prompts.

[0135] Step 4: Single-round reasoning.

[0136] Specifically, the cropped ROI image and the current version of the quality inspection prompts are input into the multimodal large model, and a single round of no-conclusion injection inference is performed. That is, no known quality inspection conclusions are input into the model, and the judgment is completed solely based on the visual features of the image and the prompt rules, and finally a structured quality inspection result is output.

[0137] Step 5: Evaluation of primary and secondary indicators.

[0138] Specifically, the results are verified according to the rule that the main indicator determines the correctness or incorrectness and the auxiliary indicator determines the cause of the error: the correctness or incorrectness judgment only compares the main indicator with the human true value, and the auxiliary indicator does not participate in the accuracy statistics; if the judgment is incorrect, the specific cause of the error (visual confusion, angle occlusion, boundary blur, etc.) is located by comparing the deviation of the auxiliary indicator and combining the model reasoning description.

[0139] Step 6: Correct Judgment: If the judgment of the current sample is correct, the AI agent retains the current version of the prompt, directly proceeds to the next image sample, and continues the cyclic verification.

[0140] Incorrect Judgment: If the judgment of the current sample is wrong, the AI agent combines the manual ground truth, the model inference description and the error cause localization to directionally update the content of the prompt, which can not only modify the judgment boundary of the auxiliary indicators and supplement the anti-confusion rules, but also adjust the constraint logic of the main indicators to generate a new version of the prompt.

[0141] Step 7: Full-volume backtesting.

[0142] Specifically, after the prompt is modified, all processed samples are re-executed for the single-round inference in step 4, instead of only verifying the single sample with error. Only when all processed samples are judged correctly, the new version of the quality inspection prompt is considered qualified.

[0143] Step 8: Difficult case handling.

[0144] Specifically, if the sample still cannot be correctly identified after multiple iterative modifications and the preset iteration upper limit is reached, a difficult case fallback mechanism is started: the difficult case sample is excluded from the current iteration round, no forced rule modification is performed for adaptation, an uncertain conclusion is output for such difficult case sample during inference, and it is automatically transmitted for manual recheck.

[0145] Step 9: Counterfactual gating.

[0146] Specifically, in order to perform security verification on the quality inspection prompt before launch, it is necessary to quantify the compliance risk of the model to text injection. By constructing a counterfactual probe set, performing two rounds of wrong label injection tests, and calculating the counterfactual compliance rate, it is verified whether the model will be guided by wrong text to abandon visual judgment. If the counterfactual compliance rate is greater than or equal to a preset threshold (e.g., 30%): it is determined that there is a text manipulation risk, and the process returns to step 4 for further optimization. If the compliance rate is less than the threshold: it is determined that the independence of visual judgment meets the standard, and the current quality inspection prompt is archived as a verified prompt.

[0147] Step 10: Deployment and inference.

[0148] Specifically, the verified prompt is officially put into production line deployment, supports two image acquisition methods of handheld mobile devices and fixed station cameras, and adapts to the detection requirements of different stations on the production line.

[0149] In one embodiment, as Figure 9 shown, a module diagram of an automatic screener for images of liquid-cooled tube connectors is provided. The automatic screener includes: a data acquisition module, a region positioning module, a region intercepting module, a multimodal large model inference module, an evaluation module, an agent module, a counterfactual gating module, and a quality inspection output module, wherein:

[0150] Data acquisition module corresponding Figure 8 Step 1, Sample and Task Configuration, serves as the data entry point for the automated image screening system for liquid cooling pipe connectors. It is responsible for acquiring raw images of the liquid cooling connectors, compatible with both fixed-station cameras on the production line and handheld mobile devices used by quality inspectors, covering all workstation scenarios including the cold plate side, the housing side, and the housing connection points. Simultaneously, it is accompanied by standardized labeled data provided by experienced quality inspectors, including the true value of the main connection status indicator, as well as auxiliary indicator labels such as groove visibility and red latch status, providing a unified true value benchmark for subsequent reasoning, evaluation, and iteration.

[0151] Corresponding area positioning module Figure 8 Step 2: This module independently calls the multimodal large model, paired with fixed, dedicated positioning prompts. It is only responsible for identifying the location of the area where the black plastic shell of the connector meets the port in the image, outputting the bounding box coordinates, and does not participate in any qualification judgment. The positioning logic is completely independent of the quality inspection rules, avoiding interference from the quality inspection rule iteration with the positioning accuracy, and also reducing the interference of irrelevant background on subsequent quality inspections.

[0152] The corresponding flowchart step 2 for the region capture module. Based on the coordinates output by the region positioning module, the original image is precisely captured to generate a sub-image containing only the ROI (Region of Interest) of the core joint area of ​​the connector. Irrelevant backgrounds such as distant hardware, cables, and labels are removed from the image, providing focused high-definition input for subsequent quality inspection inference and effectively improving the accuracy and stability of the model's judgment.

[0153] Multimodal large model inference module corresponding Figure 8 Step 4 is the core execution unit for quality inspection judgment. It receives the cropped ROI image and the current version of the quality inspection prompts, and strictly executes the single-round no-conclusion injection inference rule: the model makes independent judgments based solely on the image's visual features and the judgment rules in the prompts, without injecting any known quality inspection conclusions throughout the process. Finally, it outputs a standardized structured result containing the red latch status, groove visibility, and connectivity status, ensuring the authenticity and purely visual attributes of the judgment result.

[0154] Evaluation module corresponding Figure 8 Step 5 is the accuracy verification and error localization stage. It follows three judgment rules: primary indicators are judged independently, auxiliary features are not rejected when they are not identifiable, and auxiliary indicators only provide support, not substitution. This module only compares the connection state primary indicator with the human-generated truth value to determine the correctness of the sample; simultaneously, by comparing the identification deviation of auxiliary indicators and combining the model's reasoning description to locate the specific error cause, it filters out samples that failed the judgment and sends them to the intelligent agent module for targeted optimization.

[0155] Intelligent agent module corresponding Figure 8Steps 3, 6, 7, and 8 are used to generate the initial version of the quality inspection prompt words; receive the error samples and error causes output by the evaluation module, add, delete, and modify the description of auxiliary indicators, correct the feature judgment boundary, and supplement the anti-confusion rules; automatically perform full sample backtesting after each word modification to ensure that rule modification does not introduce new misjudgments; uniformly manage all prompt word versions, handle the removal of difficult cases after reaching the iteration limit, and realize the automated closed-loop optimization of prompt words.

[0156] Counterfactual gating module, corresponding Figure 8 Step 9 is the mandatory security verification module before the prompt words go live. This involves building a counterfactual probe set for label flipping, performing two rounds of erroneous label injection tests, calculating the model's counterfactual compliance rate, and performing gating judgments against preset thresholds. Only prompt word versions with compliance rates below the threshold and meeting the model's visual judgment independence requirements are allowed to be officially launched and archived, thus mitigating security risks.

[0157] Quality inspection output module corresponding Figure 8 Step 10. Output three standardized quality inspection conclusions for the actual production line: qualified, unqualified, and uncertain. Samples marked as uncertain are automatically transferred to the manual re-inspection station.

[0158] In one embodiment, such as Figure 10 As shown, the current prompt word Pn serves as the starting input for this iteration, and Pn represents the effective quality inspection prompt word of the nth version.

[0159] Single-round reasoning: ROI image + Pn → decision Di. Only the cropped ROI image and the current version prompt are input. The multimodal large model is called to complete one reasoning and output the structured decision result Di.

[0160] Internal comparison annotation (without injecting conclusions into the API) compares the model's output decision Di with the manually annotated true value to determine whether the inference is correct.

[0161] Correct → Next. If the comparison result matches the true value, the current Pn version remains unchanged, and the process proceeds directly to the next sample, continuing the same single-round inference and verification process.

[0162] Error → Agent modifies word to get Pn+1. If the comparison judgment result does not match the true value, the agent combines the deviation points of the main and auxiliary indicators and the model reasoning description to locate the specific error cause, and modifies the prompt word to generate an optimized new version prompt word Pn+1.

[0163] Full backtesting has processed the entire dataset. After the new version's prompts are generated, a single round of inference will be re-executed on the entire dataset of all processed samples.

[0164] If the error persists but the iteration count has not exceeded the limit, modify the word; if the limit is reached, remove the difficult case. If the backtest still contains errors, but the number of iterations has not reached the preset limit, continue to modify the word for the new error and enter the next round of optimization loop; if the iteration limit has been reached and the error still cannot be corrected, the sample is determined to be an extremely difficult case and removed from the current iteration test set. During inference, the output is uncertain and is manually reviewed to avoid compromising the universality of the overall rule in order to accommodate a small number of difficult cases.

[0165] Implement counterfactual gating. Once all samples are either correctly identified or difficult cases have been eliminated, the iteration loop officially ends, and the process enters the security verification stage before the quality inspection prompts are launched: the text compliance risk of the model is quantified through counterfactual label experiments, and only quality inspection prompts with compliance rates below the threshold can be officially archived and launched.

[0166] In one embodiment, an automatic screening method for images of liquid cooling pipe connectors is provided, comprising:

[0167] Extract the connector image containing the liquid cooling pipe connector from the image data of the liquid cooling pipe connector;

[0168] Determine the connector type of the cold-rolled tube connector based on the connector image;

[0169] The corresponding quality inspection prompt is invoked according to the connector type; the quality inspection prompt and connector image are input into the preset multimodal large model to obtain the first quality inspection result output by the multimodal large model;

[0170] The quality inspection prompts are adjusted based on the first quality inspection result to obtain the adjusted quality inspection prompts. The adjusted quality inspection prompts are used to input the preset multimodal large model to perform quality inspection and recognition on the connector image.

[0171] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods according to the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0172] It is obvious to those skilled in the art that the modules or steps of the present invention described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. They can be implemented using computer-executable program code, and thus can be stored in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those described herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the present invention is not limited to any particular combination of hardware and software.

[0173] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An automatic image screening device for liquid-cooled pipe connectors, characterized in that, include: The image acquisition module is used to send the acquired image data of the liquid cooling pipe connector to the processing module; The processing module is connected to the image acquisition module and is used to extract a connector image containing a liquid cooling pipe connector from the image data. The connector type of the liquid cooling pipe connector is determined based on the connector image; Call the corresponding quality inspection prompt word according to the connector type; The quality inspection prompts and the connector image are input into a preset multimodal large model to obtain the first quality inspection result output by the multimodal large model; The quality inspection prompts are adjusted based on the first quality inspection result to obtain the adjusted quality inspection prompts. The adjusted quality inspection prompts are used to input the preset multimodal large model to perform quality inspection recognition on the connector image.

2. The automatic image screening device for liquid-cooled pipe connectors according to claim 1, characterized in that, The processing module is further configured to determine whether the first quality inspection result is correct based on the first quality inspection result, the connector image, and the quality inspection indicators of the connector type corresponding to the connector image; and if the first quality inspection result is determined to be incorrect, to determine the cause of the error based on the first quality inspection result and the manual annotation result. Adjust the quality inspection prompt words according to the cause of the error.

3. The automatic image screening device for liquid-cooled pipe connectors according to claim 2, characterized in that, The processing module is further configured to determine that the first quality inspection result is incorrect and the reason for the error is that the quality inspection indicator was missed when the first quality inspection result is that the quality inspection passed and the manual annotation result is that the quality inspection failed.

4. The automatic image screening device for liquid-cooled pipe connectors according to claim 2, characterized in that, The processing module is further configured to determine that the first quality inspection result is incorrect and the reason for the error is misjudgment of the quality inspection indicators when the first quality inspection result is that the quality inspection failed and the manual annotation result is that the quality inspection passed.

5. The automatic image screening device for liquid-cooled pipe connectors according to claim 3, characterized in that, The processing module is used to determine the recognition results of each quality inspection indicator in the connector image output by the multimodal large model based on the current quality inspection prompt words; The identification results of each quality inspection indicator are compared with the manual annotation results of each quality inspection indicator to determine the target quality inspection indicator in which the judgment is incorrect. Determine the defects in the current quality inspection prompts based on the target quality inspection indicators; Adjust the quality inspection prompts based on the defects in the prompts.

6. The automatic image screening device for liquid-cooled pipe connectors according to claim 5, characterized in that, The defects in the prompt words include at least one of the following: visual confusion, over-judgment when features are not visible, blurred judgment boundaries, incorrect connector type identification, and judgment logic error.

7. An automatic image screening device for liquid-cooled pipe connectors according to any one of claims 1-6, characterized in that, The processing module is used to input the image data and preset recognition prompts into a multimodal large model to identify key areas in the image data that contain liquid cooling pipe connectors. The key region is extracted from the image data to obtain the connector image of the key region.

8. An automatic image screening device for liquid-cooled pipe connectors according to any one of claims 1-6, characterized in that, The processing module is used to receive image data of liquid cooling pipe connectors of various different connector types as manually labeled sample data; The system receives annotation information input by the user, annotates the sample data according to the annotation information, and obtains the manual annotation result of the sample data.

9. An automatic image screening device for liquid-cooled pipe connectors according to any one of claims 1-6, characterized in that, The processing module is used to input the test connector image into the multimodal large model, wherein the test connector image is an image whose quality inspection results have been confirmed to be correct; Input counterfactual prompts into the multimodal large model to obtain a second quality inspection result, wherein the counterfactual prompts include guiding words for quality inspection results that are contrary to the confirmed quality inspection results of the test connector image; If the second quality inspection result is inconsistent with the confirmed quality inspection result of the test connector image, the quality inspection prompt words are adjusted to increase the dependence of the multimodal large model on visual features.

10. The automatic image screening device for liquid-cooled pipe connectors according to claim 9, characterized in that, The processing module is used to input multiple different test connector images and counterfactual prompts into the multimodal large model to obtain multiple second quality inspection results output by the multimodal large model; Based on the multiple second quality inspection results and the confirmed quality inspection results of the multiple different test connector images, determine the proportion of erroneous results; If the proportion of error results is greater than or equal to the proportion threshold, the quality inspection prompts are adjusted to increase the multimodal large model's dependence on visual features, and the process returns to the step of inputting multiple different test connector images and the counterfactual prompts into the multimodal large model until the proportion of error results is less than the proportion threshold.

11. The automatic image screening device for liquid-cooled pipe connectors according to claim 10, characterized in that, The processing module is used to archive the current quality inspection prompt word as a correct prompt word when the proportion of the error result is less than the proportion threshold.

12. An automatic image screening device for liquid-cooled pipe connectors according to any one of claims 1-6, characterized in that, The processing module is used to re-input the adjusted quality inspection prompts and the connector image into a preset multimodal large model to obtain a new first quality inspection result output by the multimodal large model; If the new first quality inspection result is incorrect, adjust the adjusted quality inspection prompt based on the new first quality inspection result, and return to the step of re-inputting the adjusted quality inspection prompt and the connector image into the preset multimodal large model until the new first quality inspection result is correct.

13. The automatic image screening device for liquid-cooled pipe connectors according to claim 12, characterized in that, The processing module is used to record the number of iterations for returning to the step of re-entering the adjusted quality inspection prompt and the connector image into the preset multimodal large model; If the number of iterations reaches the preset iteration limit, stop using the multimodal large model to output the first quality inspection result; The connector image is transmitted to a manual inspection station for quality inspection by a human.

14. The automatic image screening device for liquid-cooled pipe connectors according to claim 12, characterized in that, The processing module is used to archive the current adjusted quality inspection prompt as the correct prompt if the new first quality inspection result is correct.

15. An automatic image screening device for liquid-cooled pipe connectors according to any one of claims 1-6, characterized in that, The quality inspection indicators include primary indicators and secondary indicators. The quality inspection prompts are used to instruct the multimodal large model to output quality inspection results based on the primary indicators, determine the primary indicators based on the secondary indicators, and not participate in the judgment of the primary indicators if the secondary indicators are unrecognizable.

16. An automatic image screening device for liquid-cooled pipe connectors according to any one of claims 1-6, characterized in that, The processing module is used to identify the geometry, snap-fit ​​direction, and port connection method of the liquid cooling pipe connector based on the connector image; The connector type is determined based on the geometry, snap-fit ​​direction, and port connection method of the liquid cooling pipe connector.

17. The automatic image screening device for liquid-cooled pipe connectors according to claim 16, characterized in that, The processing module is used to determine that the connector type is a push-in elbow type if it is determined that the liquid cooling pipe connector includes a cylindrical elbow housing and the cylindrical elbow housing is pushed into the cooling port short pipe.

18. The automatic image screening device for liquid-cooled pipe connectors according to claim 16, characterized in that, The processing module is used to determine that the connector type is rectangular side-clamp type if it determines that the liquid cooling pipe connector includes an angular or rectangular housing and the angular or rectangular housing surrounds the cooling port short pipe from the side with a snap fastener.

19. The automatic image screening device for liquid-cooled pipe connectors according to claim 16, characterized in that, The processing module determines that the connector type is a vertical top-clamp type if it determines that the liquid cooling pipe connector includes a cylindrical housing and the cylindrical housing is snapped into the cooling port short pipe from bottom to top.

20. An automatic screening method for images of liquid cooling pipe connectors, characterized in that, An automatic screening device for images of liquid-cooled pipe connectors according to any one of claims 1-19, comprising: Extract the connector image containing the liquid cooling pipe connector from the image data of the liquid cooling pipe connector; The connector type of the cold night tube connector is determined based on the connector image; The corresponding quality inspection prompt is invoked according to the connector type; the quality inspection prompt and the connector image are input into a preset multimodal large model to obtain the first quality inspection result output by the multimodal large model; The quality inspection prompts are adjusted based on the first quality inspection result to obtain the adjusted quality inspection prompts. The adjusted quality inspection prompts are used to input the preset multimodal large model to perform quality inspection recognition on the connector image.

Images