Image recognition and intelligent valuation system for automobile gearbox old parts

Abstract

The application discloses an image recognition and intelligent evaluation system for automobile gearbox old parts, and relates to the technical field of computer vision; the application is provided with an image acquisition module, an image processing module, an inquiry template generation module, a parts supplier response module and a model iteration module. The image acquisition module acquires old part images and checks the quality through an algorithm; the image processing module calls a multi-task convolutional neural network to extract parts categories, damage types and wear quantization indexes, and calculates a value interval in combination with a damage-recycling value mapping model; the inquiry template generation module generates a standardized template and recommends suitable parts; the parts supplier response module realizes directional pushing, quotation receiving and two-way bargaining; and the model iteration module optimizes model parameters based on transaction feedback data. Through multi-module cooperation and unique technical means, the application realizes intelligent recognition and accurate evaluation of old parts, reduces the operation threshold, improves transaction efficiency, and promotes the standardized development of old part transactions.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, specifically to an image recognition and intelligent valuation system for used automotive transmission parts. Background Technology

[0002] As a core power transmission component, the remanufacturing and recycling of used automotive transmissions is a crucial way for the auto parts industry to reduce costs, increase efficiency, and practice environmental protection. In the trading of used transmission parts, stores and remanufacturing companies must first determine the type and extent of damage of the used parts, and then assess their remanufacturability based on market conditions and repair costs. This assessment serves as the core basis for negotiating prices. The professionalism of this process directly determines the efficiency and fairness of the transaction, becoming a key link affecting the circulation of used transmission parts.

[0003] However, in current technologies, the identification and valuation of used transmission parts mainly rely on manual experience. Staff must use their accumulated professional knowledge to determine the type, type of damage, and degree of wear of the used parts through visual observation and simple tool measurements, then combine this with subjective experience to estimate their value. This method not only demands a very high level of expertise from operators, making it difficult for most small and medium-sized shops to independently achieve accurate valuations, but also suffers from inefficiency, large valuation errors, and mismatches between demand and supply. This can easily lead to transaction disputes, waste of used parts resources, and severely hinder the standardization and efficiency of used transmission parts trading.

[0004] In view of the above, this application is hereby submitted. Summary of the Invention

[0005] The purpose of this invention is to provide an image recognition and intelligent valuation system for used automotive transmission parts, in order to solve the problems mentioned in the background art.

[0006] To address the aforementioned technical problems, this invention provides an image recognition and intelligent valuation system for used automotive transmission parts, comprising an image acquisition module, an image processing module, a price inquiry template generation module, a parts supplier response module, and a model iteration module. The image acquisition module acquires image data of the used transmission parts. The image processing module utilizes a multi-task convolutional neural network to extract the part category, damage area, damage type, and wear quantification index of the used transmission parts. Furthermore, it calculates the remanufacturable value range of the used transmission parts by constructing a damage-remanufacturing value mapping model. The mapping relationship of the damage-remanufacturing value mapping model is as follows:

[0007] , in The midpoint of the remanufacturable value range. The basic market value for undamaged used transmission parts. For the first The weighting coefficients of class value influencing factors, , For the first The system quantifies the four categories of value-influencing factors: damage parameter factor, component function weight factor, repair process cost factor, and market suitability factor. The inquiry template generation module generates standardized inquiry templates based on the image processing module's output and recommends matching new or remanufactured parts. The parts supplier response module pushes the standardized inquiry templates to suitable parts suppliers, receives price quotes from suppliers, and supports two-way negotiation between stores and parts suppliers. The model iteration module collects transaction prices from actual transactions, damage parameter deviations from physical verification feedback, and repair effect data, and applies the feature extraction weights of the multi-task convolutional neural network to the weight coefficients in the damage-remanufacturing value mapping model. Dynamic adjustments are made; through a unique damage-remanufacturing value mapping model and integration with the entire process module, a closed loop is achieved from old part image collection to value assessment, transaction matching and model optimization, solving the problems of traditional manual valuation relying on experience, large deviations and low efficiency. At the same time, through two-way negotiation and model iteration functions, the system's practicality and long-term adaptability are further improved, and the professional threshold for stores is significantly reduced.

[0008] Furthermore, the image acquisition module includes an image acquisition unit and an acquisition auxiliary unit. The image acquisition unit is used to capture multi-view images of the front panoramic view, side panoramic view, and close-up of the suspected damaged area of ​​the old gearbox part through a mobile terminal, and is accompanied by a solid color background and scale reference piece during shooting. The acquisition auxiliary unit analyzes the sharpness, illumination uniformity, and integrity of the old part in real time through a sharpness verification algorithm, an illumination uniformity analysis algorithm, and an old part integrity verification algorithm, and outputs corresponding adjustment prompts. By specifying the specific perspective of image acquisition, auxiliary tools, and verification algorithms, it ensures that the acquired image data meets the accuracy requirements of subsequent recognition and analysis, reduces recognition errors caused by image quality issues, and provides a reliable data foundation for subsequent accurate feature extraction and value calculation.

[0009] Furthermore, the image processing module includes an image preprocessing submodule, a feature extraction submodule, and a value calculation submodule. The image preprocessing submodule performs Gaussian filtering for noise reduction, adaptive threshold segmentation combined with morphological opening operations for reflective oil stain suppression, histogram equalization enhancement, and contour correction on the image data acquired by the image acquisition module. The feature extraction submodule outputs the accessory category, damage type, location information of the damaged area, and wear quantification index through a multi-task convolutional neural network. Based on the output of the feature extraction submodule, the value calculation submodule calculates the remanufacturable value range through a damage-remanufacturing value mapping model. When the repair cost is greater than 1.2 times the median value of the remanufacturable value range, the old part is determined to be unremanufacturable, and the value range is output as 0. By clarifying the specific algorithm of image preprocessing, the complete output content of feature extraction, and the logic for determining remanufacturing feasibility, the image processing flow becomes more logical and operable, improving the recognition of image features and the scientific nature of value calculation, further ensuring the accuracy of system identification and valuation.

[0010] Furthermore, the damage types output by the feature extraction submodule include cracks, wear, fatigue spalling, deformation, and oil passage blockage. The wear quantification indicators include crack length, crack width, wear depth, fatigue spalling area, and the proportion of oil passage blockage area. The specific range of damage types and wear quantification indicators is clearly defined, comprehensively covering the core damage characteristics of old gearbox parts. This provides rich and crucial input parameters for the damage-remanufacturing value mapping model, ensuring the comprehensiveness and accuracy of remanufacturable value calculation.

[0011] Furthermore, the inquiry template generation module includes a template construction unit and a matching recommendation unit. The template construction unit integrates parts category, damage type, damage area, wear quantification index, and remanufacturable value range to generate a standardized inquiry template that includes parts specifications, remanufacturing requirement level, quotation validity period, and logistics requirements. The matching recommendation unit, based on parts category, remanufacturing requirement level, and compatible gearbox model, uses a cosine similarity algorithm to filter and recommend the new or remanufactured parts with the highest compatibility from a preset parts database, prioritizing remanufactured parts. When the remanufacturing requirement level is complex repair, the corresponding new parts are recommended simultaneously. The complete fields of the inquiry template and the matching recommendation algorithm and rules are refined. The standardized inquiry template ensures the standardized transmission of transaction information, and the accurate matching recommendation function reduces demand mismatch between parts suppliers and stores, improving the relevance of quotations and the success rate of transactions.

[0012] Furthermore, the parts supplier response module includes a template push unit, a quotation receiving unit, and a negotiation unit. The template push unit pushes standardized inquiry templates to suitable parts suppliers based on the parts supplier's business scope, repair capabilities, and historical transaction data. The quotation receiving unit receives and stores quotation information submitted by parts suppliers for recommended new or remanufactured parts, including quotation amount, warranty period, and delivery cycle. The negotiation unit allows stores to initiate negotiation requests to parts suppliers and submit desired quotations, receive negotiation feedback from parts suppliers, and display it to the stores. By breaking down the complete functional units of the parts supplier response module, the basis for targeted push, the specific content of quotation information, and the two-way negotiation process are clarified, ensuring that the inquiry template accurately reaches suitable parts suppliers, standardizing quotation information management, and promoting reasonable transaction prices between both parties through the negotiation function, thereby reducing transaction disputes.

[0013] Furthermore, the model iteration module includes a data collection unit, a deviation analysis unit, and a parameter adjustment unit. The data collection unit is used to collect actual transaction prices, deviation values ​​of damage parameters between physical objects and image recognition, quality inspection data of repaired old parts, and transaction evaluation feedback data. The deviation analysis unit is used to statistically analyze the deviation rate of damage parameter recognition and the deviation rate of value assessment, and to filter samples with deviation rates exceeding a preset threshold. The parameter adjustment unit is used to adjust the feature extraction weights and weight coefficients in the damage-remanufacturing value mapping model of the multi-task convolutional neural network using an incremental training method based on the filtered samples. Updates and adjustments were made; the complete functional units and workflow of the model iteration module were refined. Through precise data collection, targeted deviation analysis, and scientific parameter adjustment methods, the effectiveness and relevance of model iteration were ensured, enabling the system to continuously optimize performance and adapt to new damage types and market changes.

[0014] Furthermore, the parts database stores various new and remanufactured parts, including model specifications, compatible transmission types, manufacturer information, repair process levels, and historical price data. The matching and recommendation unit compares and filters the data in the parts database based on the parts category, remanufacturing requirement level, and compatible transmission model, outputting the top 10 recommendations with the highest compatibility and providing reasons for the recommendations. This clarifies the complete storage content of the parts database and the rules for outputting recommendation results, improving the accuracy and transparency of the recommendations. This allows stores to quickly obtain information on new or remanufactured parts that meet their needs, while also providing a reference for parts suppliers to quote prices, further promoting efficient transaction completion.

[0015] Furthermore, the quantification method for the four types of value-influencing factors is as follows: damage parameter factor. ,in , , , , The weighting coefficients for cracks, wear, fatigue spalling, oil passage blockage, and deformation are respectively, and their sum is 1. , , , , These are the normalized values ​​of the quantitative indicators corresponding to the damage; and the component functional weighting factor. Preset weighting coefficients for core or secondary functional components; repair process cost factors. , The cost of the repair process corresponding to the damage, The market price of this type of new part; market fit factor: , For adjustment coefficients, For the supply and demand ratio of used parts, The average supply-demand ratio for this type of used parts is defined; the specific quantitative formulas and parameter meanings of the four value influencing factors are clarified, making the calculation logic of the damage-remanufacturing value mapping model clearer and more reproducible, avoiding value assessment bias caused by the ambiguity of factor quantification, and further improving the scientificity and credibility of the valuation results.

[0016] Compared with the prior art, the beneficial effects of the present invention are: 1. By constructing a complete technical system with multi-module collaboration, the pain points of traditional methods are comprehensively addressed. The image acquisition stage employs a dedicated correlation mechanism between the type of used parts and shooting parameters, combined with multi-dimensional quality verification and standardized auxiliary tools, effectively solving the challenges of capturing reflections, oil stains, and minor damage, providing a high-quality data foundation for subsequent identification. The image processing stage designs a multi-task convolutional neural network adapted to used gearbox parts, enabling the simultaneous extraction of part categories, damage types, and quantitative indicators, breaking through the limitation of general identification which can only provide qualitative descriptions, and providing a scientific basis for value assessment.

[0017] 2. In the value assessment stage, a four-dimensional value influencing factor system is constructed. Combined with an improved gradient boosting tree algorithm, a non-linear mapping model is established to achieve full-dimensional value transformation from damage parameters and component functions to market adaptability. This significantly improves the accuracy and consistency of valuation and reduces transaction disputes. In the inquiry and transaction stages, standardized templates, cosine similarity matching, and targeted push notifications based on component supplier capability profiles are used. Two-way negotiation and electronic contract mechanisms are introduced to standardize the transaction process, reduce ineffective communication, and improve transaction efficiency and standardization.

[0018] 3. In the model iteration stage, a data closed-loop optimization system based on transaction feedback is constructed. Through precise data collection, targeted deviation analysis, and incremental training, the model is dynamically optimized, enabling the system to continuously adapt to new damage types and market changes, maintain high identification accuracy and valuation accuracy in the long term, extend the system's life cycle and core competitiveness, and promote the standardized and efficient development of old gearbox parts recycling. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the principle of an image recognition and intelligent valuation system for used automotive transmission parts. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please see Figure 1 This invention provides a technical solution: an image recognition and intelligent valuation system for used automotive transmission parts. This embodiment details the entire implementation process from image acquisition, image processing, value assessment, inquiry template generation, parts supplier response, to model iteration. The system uses computer vision technology as its core, combining a multi-task convolutional neural network and a damage-remanufacturing value mapping model to achieve automatic identification, accurate valuation, and efficient transaction matching of used transmission parts, solving the technical problems of high professional threshold, low efficiency, and large deviations in traditional manual valuation.

[0022] (I) Image Acquisition Steps: Image acquisition is the foundation for subsequent system recognition and valuation. The core objective is to obtain clear, complete, and interference-free images of used gearbox parts to ensure the accuracy of subsequent feature extraction. Traditional image acquisition relies on manual, arbitrary shooting, which is prone to recognition errors due to improper angles, lighting interference, and missing details. Therefore, this step ensures image data quality through standardized acquisition specifications, clear auxiliary tool configurations, and multi-dimensional quality verification mechanisms. The core algorithms used in this step include the Laplacian operator sharpness verification algorithm, the grayscale histogram illumination uniformity analysis algorithm, and the contour matching integrity verification algorithm. These are used in conjunction with a mobile app to guide acquisition and provide real-time quality control. All algorithm parameters are clearly defined and adjustable to meet the needs of different usage scenarios. Specific technical means are as follows: (1) Preparation before data collection: Hardware configuration: Use a mobile device that supports high-definition shooting, with a camera resolution of no less than 1080P and autofocus to ensure clear capture of minor damage to old parts; equip a solid-color background board in neutral gray to eliminate interference from complex backgrounds on image recognition; configure a scale reference piece with uniform scale lines spaced 10 mm apart on the surface, made of high-contrast black material, for calibration of the size of old parts in subsequent images. The scale reference piece should be no less than 20 cm long and no less than 5 cm wide to ensure accurate size calibration. Software configuration: Install a dedicated data collection APP on the mobile device. The APP has built-in image quality verification module, data collection guidance module, old part type selection module, and image storage and transmission module, and supports iOS 12.0 and above, and Android 8.0 and above operating systems to ensure compatibility.

[0023] (2) Data Acquisition Operation Process: Old Part Type Selection: After opening the APP, the user selects the specific type of old part to be acquired from the preset list of old gearbox parts. This list covers core gearbox components such as valve body, clutch plate, planetary gear set, input shaft, output shaft, housing, synchronizer, and bearing. Each component is equipped with a corresponding schematic diagram for easy identification and selection by the user. The selection result will serve as the initial reference information for subsequent image processing and will be used to match the exclusive recognition model parameters. Shooting Parameter Settings: The APP automatically matches the preset shooting parameters according to the old part type selected by the user. For metal parts with strong surface reflection (such as valve body and gear), the exposure is automatically reduced by 15%-20%, and the white balance is adjusted to cloudy mode to avoid reflection interference; for darker parts (such as housing), the exposure is automatically increased by 10%-15%, and the white balance is adjusted to automatic mode to ensure uniform image brightness; the shooting focus mode for all components is set to continuous autofocus to ensure image clarity. Multi-view shooting: Following the app's guidance, users take photos of the old part from fixed angles, including three core perspectives: a front panoramic view, a side panoramic view, and a close-up of the suspected damage area. The front panoramic view is used to identify the overall outline and category of the old part; the old part must be completely centered in the frame with no cropping at the edges. The side panoramic view is used to determine the three-dimensional structure and assembly relationship of the old part; the shooting angle is 90 degrees from the front. The close-up of the suspected damage area is used to capture details such as tiny cracks and wear; the suspected damage area must fill the frame, with a magnification of at least 2x. Image quality verification: After each image is taken, the app immediately initiates a quality verification process. If the verification fails, the user is prompted to retake the photo until a satisfactory image is obtained. Sharpness verification: The Laplacian operator is used to calculate the image sharpness value. The sharpness value calculation formula is: ,in This is the sharpness value. For the image in coordinates The pixel grayscale value at that location, For the Laplace operator, , and These are the width and height in pixels, respectively. Preset sharpness threshold. ,when If the image is deemed blurry, the app will prompt the user, "The image is blurry; please keep your phone stable and take the picture at a distance of 30-50 cm from the original object." Illumination uniformity analysis: Illumination uniformity is analyzed by calculating the gray-level histogram distribution of the image. Illumination uniformity coefficient. The calculation formula is: , in This represents the maximum frequency in the grayscale histogram. This represents the minimum frequency in the grayscale histogram. This represents the average frequency of the grayscale histogram. Preset illumination uniformity threshold. ,when If the image is unevenly lit, indicating areas that are too bright or too dark, the app will prompt the user with "Uneven lighting, please adjust the shooting angle to avoid direct sunlight or shadows." For old item integrity verification: The app extracts the old item region from the captured image using an image segmentation algorithm. The segmentation algorithm combines threshold segmentation and edge detection. First, Otsu threshold segmentation is used to initially extract the foreground region, and then Canny edge detection is used to optimize the contour. The extracted old item region is compared with the preset standard contour for this type of old item to calculate the contour matching degree. , , in The area of ​​overlap between the extracted old part region and the standard contour. Area of ​​the standard contour. Preset contour matching threshold. ,when If the app determines that the old part was not fully captured, it will prompt the user with "The old part is not fully presented. Please capture the missing part in the photo." The missing part is the key area.

[0024] (3) Image storage and transmission: The images that have passed the quality check are stored locally by the APP according to the naming rule of "old type-shooting time-viewpoint". The storage path is the APP's exclusive folder to avoid accidental deletion. At the same time, the image data is encrypted using the AES-256 encryption algorithm and then transmitted to the system cloud server via the HTTPS protocol. The storage format is JPG and the compression quality factor is set to 0.8 to ensure the security, readability and transmission efficiency of the image data.

[0025] Example: After disassembling a Volkswagen DSG 7-speed dual-clutch transmission, a car repair shop needs to assess the value of the faulty valve body. The shop staff opens a data collection app, selects "dual-clutch valve body" from the list of used parts, and the app automatically configures the shooting parameters: exposure reduced by 18%, white balance set to cloudy mode, and focus mode set to continuous autofocus. The staff places the valve body on a neutral gray background, places a scale reference piece next to the valve body, ensuring the scale reference piece is on the same plane as the valve body and unobstructed, and then, following the app's guidance, takes a series of photos: a full-view shot of the valve body's front, a full-view shot of the side, and a close-up of the suspected cracked area near the oil passages.

[0026] After the photo is taken, the app checks the sharpness of the frontal panoramic image and calculates its sharpness value. , The clarity meets the requirements; the illumination uniformity coefficient... , Uniform illumination; the valve body region is extracted using an image segmentation algorithm, and the contour matching degree is obtained by comparing it with a standard contour. , The valve body was photographed completely. All three images passed verification, were successfully stored locally, and then encrypted before being transmitted to the cloud server.

[0027] The unique technical approach of this solution lies in its scene-specific association and matching of old part types with shooting parameters. It achieves real-time control of the acquisition process through multi-dimensional image quality verification algorithms, clearly defining the specific specifications of auxiliary tools and shooting angle requirements, rather than relying on traditional indiscriminate shooting. Compared to existing technologies that solely rely on manual shooting experience, this step, through standardized processes, clear tool configurations, and algorithm verification, significantly reduces the proportion of unqualified images such as blurry images, uneven lighting, and missing details, improving the accuracy of subsequent feature extraction by over 30%. Simultaneously, it lowers the operational difficulty for users, allowing them to acquire qualified images without professional photography knowledge, reducing the operational threshold by over 50% and providing a high-quality data foundation for subsequent image processing.

[0028] (II) Image Processing Steps: Image processing is the core of the system, aiming to extract the component category, damage type, damage area, and wear quantification indicators of the used transmission parts from the acquired images, providing data support for subsequent value assessment. Traditional image processing often uses single feature extraction algorithms, which struggle to simultaneously and accurately identify component categories and complex damage, and cannot achieve quantitative damage analysis, thus failing to meet the needs of used transmission part valuation scenarios. This step employs a multi-task convolutional neural network (MT-CNN) adapted for used transmission parts, combined with scenario-specific image preprocessing algorithms and damage quantification models, to achieve simultaneous extraction and accurate quantification of multi-dimensional information. All technical methods are optimized for the metal material and specific damage types of the used transmission parts. Specific technical methods are as follows: (1) Image preprocessing: The purpose of image preprocessing is to eliminate noise, reflection, oil stains and other interference factors in the image, enhance the contrast between damage features and background, and lay the foundation for feature extraction. All preprocessing algorithms are designed for the material characteristics and interference types of old gearbox parts: Noise removal: Gaussian filtering algorithm is used to denoise the image. The convolution kernel formula of Gaussian filtering is: , in The standard deviation of the Gaussian kernel, set to 1.2, is used to balance noise reduction and feature preservation. This algorithm smooths random noise in the image while preserving edge information of damaged features. Reflection and oil stain suppression: To address the common reflection and oil stain interference on the surface of old gearbox parts, an adaptive threshold segmentation algorithm combined with morphological opening operation is used. First, the image is divided into reflective, oil stain, and normal regions using the adaptive threshold segmentation algorithm. The segmentation threshold is dynamically calculated based on the local pixel gray-level mean. Then, a multi-scale Retinex enhancement algorithm is used to reduce brightness in the reflective region, with a scale parameter set to 3 and an enhancement coefficient set to 0.8. Morphological opening operation is used to remove gray-level interference in the oil stain region, with the formula: ,in For the original image, It is a 3×3 rectangular structural element. Indicates corrosion operation. This indicates a dilation operation, which removes low-grayscale interference from oil stains through a process of erosion followed by dilation. Image enhancement: Histogram equalization is used to improve image contrast, making damage features clearer. The transformation function for histogram equalization is: , in The gray levels of the original image. This represents the total number of gray levels, with a value of 256. and The pixel size of the image. grayscale The number of pixels, The gray level is equalized. This transformation makes the gray level distribution of the image more uniform and significantly improves the distinction between damage features and the background.

[0029] (2) Training and Deployment of Multi-Task Convolutional Neural Network (MT-CNN): Network Architecture Design: MT-CNN consists of three core network layers: Candidate Region Generation Network (P-Net), Region Refinement Network (R-Net), and Output Network (O-Net). Each network layer is optimized for the features of old gearbox parts: P-Net: Uses 3×3 convolutional kernels with a stride of 1 and 16 output channels. It is used to quickly filter possible old parts regions in the image, eliminate interference from the background plate and scale reference parts, and output the coordinates and confidence of the candidate regions. R-Net: Uses 3×3 convolutional kernels with a stride of 1 and 32 output channels. It accurately locates the candidate regions, corrects edge deviations through bounding box regression, and outputs the corrected candidate region coordinates. O-Net: Employs a 5×5 convolutional kernel with a stride of 1 and 64 output channels. It serves as a multi-task output layer, simultaneously performing four tasks: parts category classification, damage type identification, damage region localization, and preliminary wear level determination. Different activation functions are used in the output layer to adapt to the requirements of each task: Softmax activation for classification and Sigmoid activation for localization. Training dataset construction: 500,000 images of old transmission parts from different brands (Volkswagen, Toyota, GM, Honda, etc.), with different ages (1-10 years) and different damage types were collected and divided into a training set (350,000 images), a validation set (100,000 images), and a test set (50,000 images) in a 7:2:1 ratio. Each sample was independently labeled by two engineers with over 5 years of transmission repair experience. The labeling included part category tags (30+ core component categories), damage type tags (cracks, wear, fatigue spalling, deformation, oil passage blockage), damage area coordinates (boundary box coordinates), and wear level tags (mild, moderate, severe, extremely severe, scrap). In case of discrepancies, a third-party senior engineer arbitrated the issue to ensure the accuracy of the labeled data. Model training: The Adam optimizer was used to train the model, with an initial learning rate of 0.001, decreasing by 10% every 10 epochs, for a total of 50 epochs. The loss function used was a multi-task loss function, with the following formula: ,in For category classification loss, the cross-entropy loss function is used; The mean squared error loss function is used to determine the loss for locating the damaged area. For damage type identification loss, the cross-entropy loss function is used; The cross-entropy loss function is used to determine the wear level. , , , The weighting coefficients are 0.3, 0.2, 0.3, and 0.2, respectively, satisfying the following conditions: This is used to balance the training priorities of the four tasks. Model deployment: The trained MT-CNN model is lightweighted by using depthwise separable convolution technology to reduce the number of model parameters, compressing the number of model parameters from over 8 million to 1.5 million, ensuring that the model can run efficiently on the cloud server, with the inference time for a single image controlled within 300 milliseconds, meeting the requirements of real-time processing.

[0030] (3) Feature Extraction and Damage Quantization: Part Category Extraction: The O-Net category classification branch outputs the specific category of the old part. This branch contains 30+ category output nodes, corresponding to 30+ core components of the gearbox. The Softmax function is used for classification, and the probability value of each category is output. The category with the highest probability value is selected as the final part category result, and the classification accuracy is not less than 97%. Damage Type Recognition: The O-Net damage type recognition branch contains 5 output nodes, corresponding to 5 exclusive damage types of the old gearbox parts. The Softmax function is used to output the probability value of each type of damage. When the probability value of a certain type of damage is greater than 0.5, it is determined that the old part has this type of damage. It supports the simultaneous recognition of multiple damage types. Damage Region Extraction: The O-Net damage region localization branch outputs the bounding box coordinates of the damage region ( ),in( ) represents the coordinates of the top-left corner of the bounding box. The coordinates of the lower right corner of the bounding box are used. A bounding box regression algorithm is employed to correct the position of the damaged area, ensuring that the positioning error of the damaged area does not exceed 5 pixels. Simultaneously, a semantic segmentation algorithm is used to segment the damaged area at the pixel level, achieving a segmentation accuracy of no less than 95%, distinguishing different types of damaged areas. Wear quantification index calculation: Based on the segmentation results of the damaged area and the dimensional calibration of the scale reference, wear quantification indices are calculated. All quantification indices are converted to actual physical units: crack length. The centerline of the crack is extracted using a skeleton extraction algorithm, and its actual length is calculated using a pixel-to-millimeter mapping relationship. The formula is as follows: ,in The pixel length of the center line. The pixel-to-millimeter conversion factor is calculated from the scale spacing of the scale reference and the corresponding pixel spacing. , The actual scale spacing (10 mm) is for the scale reference piece. This represents the pixel spacing corresponding to the scale interval. Crack width. Ten measurement points are evenly selected along the direction perpendicular to the crack centerline. The crack pixel width at each measurement point is calculated, and the average value is then converted using a pixel-to-millimeter conversion factor. To convert to actual width, use the following formula: , in For the first Crack pixel width at each measurement point. Wear depth. By analyzing the grayscale difference between the worn area and the normal area, and combining this with the reflective properties of the metal parts, a mapping model between grayscale difference and wear depth is established. The formula is as follows: ,in The average grayscale difference between the worn area and the normal area. and These are the model parameters, obtained by fitting experimental data. , The unit is millimeters. Fatigue spalling area. : Calculate the pixel area of ​​the fatigue spalling region Combined with pixel-to-millimeter conversion factor To obtain the actual area Percentage of oil passage blockage area For components with oil passages (such as valve bodies), the pixel area of ​​the total oil passage region is extracted using a semantic segmentation algorithm. pixel area of ​​the blocked region Calculate the proportion .

[0031] (4) Remanufacturing Feasibility Assessment: Based on wear quantification indicators and repair cost assessment, the remanufacturing feasibility of the old parts is determined. Repair Cost The cost is obtained by querying a pre-set repair process cost database, which stores the repair processes and unit costs corresponding to various damage types. , ( If the value is the midpoint of the value assessment range, or if the core functional component suffers irreparable damage such as breakage or large-area peeling, the old part is deemed unmanufacturable, and the value range output is 0.

[0032] Example: Following the image of the Volkswagen DSG 7-speed dual-clutch valve body acquired in the previous step, the cloud server receives the image and initiates the preprocessing process. First, it uses... The Gaussian filtering algorithm was used to remove image noise. Then, an adaptive threshold segmentation algorithm was used to identify three reflective areas and one oily area on the valve body surface. Multi-scale Retinex enhancement was performed on the reflective areas (scale parameter 3, enhancement coefficient 0.8), and morphological opening operation (3×3 rectangular structuring element) was performed on the oily area. Finally, histogram equalization was used to improve image contrast.

[0033] The preprocessed image is input into the lightweight MT-CNN model. P-Net quickly filters out the valve body region, eliminating interference from the background plate and scale reference, and outputs candidate regions with a confidence level of 0.98. R-Net accurately locates the candidate regions, corrects edge deviations, and outputs bounding box coordinates (120, 150, 480, 600). O-Net outputs the accessory category as dual-clutch valve body (probability 0.99), the damage type as crack (probability 0.97), and the wear level as moderate wear (probability 0.92).

[0034] Calculate pixel-to-millimeter conversion factor based on scale reference. The 10mm scale on the reference piece corresponds to a pixel pitch of 100 pixels, therefore Millimeters per pixel. The pixel length of the crack is obtained using a skeleton extraction algorithm. pixels, then crack length Millimeters; Ten measurement points were taken perpendicular to the crack centerline, with pixel widths of 5, 6, 5, 4, 5, 6, 5, 4, 5, and 6 pixels respectively, and an average value of 5 pixels. Crack width Millimeters; obtained through grayscale difference analysis Calculate the wear depth Millimeters; valve body oil passages are unblocked. No fatigue spalling or deformation damage. millimeters 2 , .

[0035] Search the repair process cost database to find the laser cladding repair process cost for dual-clutch valve body cracks. Yuan, subsequent value assessment obtained Yuan, The part was determined to be remanufacturable. The final feature extraction results are as follows: Part category: dual-clutch valve body; Damage type: crack; Crack length: 4.5 mm; Crack width: 0.5 mm; Wear depth: 0.3 mm; No fatigue spalling, deformation, or oil passage blockage.

[0036] The unique technical approach of this solution lies in its design of a multi-task convolutional neural network architecture adapted to used transmission parts. This architecture integrates part category classification, damage type identification, damage area localization, and wear level determination into a unified training and inference process. Simultaneously, it constructs a quantitative calculation model specific to the damage types of used transmission parts, rather than employing general image recognition algorithms. Compared to existing technologies that rely on single-task identification and qualitative damage description, this step achieves simultaneous extraction and precise quantification of multi-dimensional information. The accuracy rate for identifying minute damage (≤0.5 mm) is increased to over 95%, and the error of wear quantification indicators is controlled within 5%. It not only accurately identifies the type of used part and damage type but also outputs specific quantitative indicators, providing scientific data support for subsequent value assessment. This significantly improves the practicality and accuracy of the identification results and reduces valuation bias caused by insufficient information.

[0037] (III) Value Assessment Steps: The core objective of the value assessment step is to calculate the remanufacturable value range of used transmission parts based on the part category, damage type, damage area, and wear quantification indicators obtained from image processing, providing a core basis for generating the quotation template. Traditional used part valuation relies on manual judgment combined with experience and market conditions, which is highly subjective, prone to large deviations, and cannot achieve standardized calculations, failing to meet the precise pricing requirements of used transmission parts transactions. This step constructs a four-dimensional value influencing factor system specifically for used transmission parts and uses an improved Gradient Boosting Tree (GBDT) algorithm to establish a nonlinear damage-remanufacturing value mapping model, achieving precise quantification of value. All factors and models are logically designed specifically for the value composition of used transmission parts, possessing strong scenario adaptability. Specific technical means are as follows: (1) Quantification of value impact factors: Four types of value impact factors were constructed, namely damage parameter factors. Component function weight factor Repair process cost factor and market adaptation factors The quantification methods for each factor are clear and repeatable: damage parameter factors Based on the wear quantification index obtained from image processing, the weight of each quantification index is determined using the analytic hierarchy process (AHP) to calculate the overall damage degree. The formula is as follows: ,in , , , , The weighting coefficients for cracks, wear, fatigue spalling, oil passage blockage, and deformation are 0.4, 0.3, 0.1, 0.1, and 0.1, respectively, satisfying the following conditions: ; This represents the comprehensive quantitative value of cracks. , This represents the maximum length of cracks in this type of used component. This represents the maximum width of cracks in this type of used component; This is the normalized value for wear depth. , This represents the maximum wear depth for this type of used part; This represents the normalized value of fatigue spalling area. , This represents the limit area of ​​fatigue spalling in this type of used part; This represents the normalized value of the proportion of oil passage blockage area. , This represents the maximum percentage of oil passage blockage in this type of used component; This is the normalized value of the deformation. , For deformation quantization values, This represents the limit of deformation for this type of old part. Component functional weighting factor. The weighting levels are determined based on the functional importance of the old transmission components, with weight coefficients assigned to core functional components (such as valve bodies, input shafts, and planetary gear sets). Weighting coefficients for secondary functional components (such as housings, fixing bolts, and bearings) , The corresponding weighting coefficients are directly used, and the functional importance level of components is determined by consulting the gearbox structural design manual to ensure the scientific nature of the classification. Repair process cost factor. Based on the accessory category and damage type, a pre-defined repair process cost library is matched to obtain the corresponding repair process cost for that damage. After normalization, we get The formula is: ,in The market price for this type of new part is obtained by connecting to real-time price data from major auto parts e-commerce platforms, and updated at least once every 24 hours. Market adaptability factor. Based on the part category and transmission model, it connects to a real-time auto parts market database to obtain the supply and demand ratio for the used part. Supply and demand ratio After normalization, we get The formula is: , in The adjustment coefficient is set to 2. The average supply-demand ratio for this type of used parts is calculated using market data from the past three months. The supply-demand ratio is mapped to the [0,1] interval using the Sigmoid function to ensure the stability of factor quantification.

[0038] (2) Construction and Calculation of Damage-Remanufacturing Value Mapping Model: Model Construction: A nonlinear mapping model is constructed using an improved gradient boosting tree (GBDT) algorithm. Four types of value influencing factors are used as inputs, and the median value of the remanufacturable value range is used as the output. The model contains 100 decision trees, with a maximum depth of 5 for each decision tree, a minimum number of samples per leaf node of 10, and a learning rate of 0.01. The model training data uses 3+ years of historical transaction data of used gearbox parts, including 100,000+ valid transaction records. Each record contains the quantified values ​​of the four types of value influencing factors and the actual transaction price. The data comes from 30+ mainstream auto parts trading platforms and remanufacturing enterprises nationwide. Value Calculation: First, the basic market recycling value of the used part under undamaged conditions is determined. , The value deduction rate was obtained by querying the average transaction price of undamaged used items in historical transaction data, taking the average of the transaction data over the past 6 months to ensure data timeliness. Then, the value deduction rate was calculated using a mapping model. , , in For the first The weight coefficients of the class value influence factors are obtained through GBDT model training, satisfying the following conditions: The weight coefficients obtained during the initial training are , , , The median value of the ultimately remanufacturable value range. The value range is ,in This is the market volatility coefficient, which is dynamically adjusted according to the current market supply and demand situation. The value range is 0.05-0.1. When the supply-demand ratio is greater than 2, it is 0.1; when the supply-demand ratio is less than 1, it is 0.05; and when the supply-demand ratio is between 1 and 2, it is 0.08.

[0039] Example: Following the feature extraction results of the Volkswagen DSG 7-speed dual-clutch valve body obtained in the previous step, a value assessment is performed. First, the parameters and quantification values ​​of various value-influencing factors are determined: Damage parameter factors. The limiting length of this type of valve body crack millimeters, maximum width Millimeters, wear limit depth Millimeters, fatigue spalling limit area millimeters 2 Oil passage blockage limit percentage Deformation limit millimeters; weighting coefficient , , , , Calculations yielded , , , , ,but Component function weighting factor The dual-clutch valve body is the core functional component. Repair process cost factor The repair process for the valve body crack is laser cladding, and the repair cost is... Yuan, the market price of this type of new part Yuan, then Market fit factor : Check the real-time market database to see the supply and demand ratio of this model of dual-clutch valve body. Average supply-demand ratio adjustment coefficient ,but .

[0040] Calculate the value deduction rate using the GBDT model. According to historical transaction data, the average transaction price of this model of dual-clutch valve body without damage in the past 6 months is... Yuan, current supply-demand ratio Market volatility coefficient The median value Yuan, value range is Yuan. Repair cost: 800 yuan. The old part was determined to be remanufacturable, and the final output remanufacturable value range was 890-1044 yuan.

[0041] The unique technical approach of this solution lies in constructing a four-dimensional value influencing factor system that includes damage parameters, component function, repair process cost, and market adaptability. An improved GBDT algorithm is used to establish a nonlinear value mapping model, achieving a precise conversion from multi-dimensional factors to remanufacturable value, rather than using traditional single-factor or linear valuation models. Compared to existing valuation methods that rely on human experience or simple market comparisons, this step, through scientific factor quantification and nonlinear modeling, reduces the value assessment deviation rate to within 5%, a significant improvement over traditional manual valuation (deviation 20%-30%). This makes the valuation results more valuable, reduces transaction disputes caused by subjective judgment, and provides clear value evidence for stores and parts suppliers, promoting efficient transaction completion.

[0042] (IV) Inquiry Template Generation Step: The purpose of the inquiry template generation step is to generate standardized inquiry templates based on the feature results of image processing and the value range of value assessment, and to recommend matching new or remanufactured parts, providing a standardized and efficient information carrier for transaction communication between stores and parts suppliers. In the traditional inquiry process, stores need to manually write inquiry information, which suffers from problems such as incomplete information, inconsistent format, and vague expression of needs, resulting in parts suppliers' quotations being not targeted and communication efficiency being low. This step uses automated template generation technology and a parts matching recommendation algorithm based on cosine similarity to achieve standardized integration and accurate matching recommendation of inquiry information. All template fields and matching rules are designed for the information needs of used gearbox parts transactions, and have strong practicality. The specific technical means are as follows: (1) Construction of Standardized Inquiry Template: Template Field Design: The template includes five core modules: basic information of old parts, damage details, value range, remanufacturing demand level, and expected quotation requirements. All fields are automatically filled without manual intervention: Basic information of old parts module: Automatically fills in the part category, gearbox model compatibility range, service life of old parts, and material of old parts. The gearbox model compatibility range is determined by associating the part category with the brand. The service life of old parts is estimated by the wear degree of old parts in the image and the brand's manufacturing information. The material of old parts is the preset material of the corresponding component (e.g., the valve body is aluminum alloy). Damage details module: Automatically fills in the damage type, damage quantification index, and damage area description, directly referencing the output results of the image processing steps. The damage area description is converted into a text description based on the damage area coordinates (e.g., "area near the oil passage"). Value range module: Fills in the remanufacturable value range output by the value assessment steps. Remanufacturing demand level module: Based on the value deduction rate Classify remanufacturing demand levels. Level 1 (simple repair) Level 2 (routine repair) Level 3 (Complex Repair). Expected Quotation Requirements Module: Includes quotation validity period (default 7 days, modifiable), payment method (preset options: cash on delivery, 30% prepayment with final payment upon delivery), and logistics requirements (preset options: SF Express, JD Logistics). Users can select or modify according to their actual needs. Template Generation Algorithm: Employs automatic document generation technology based on XML template format. It integrates the content of each module into a standardized document according to preset format specifications, supporting PDF and Word export. It also generates digital templates that can be directly pushed into the system. The template encoding uses a unique coding rule of "store number-date-old item type" for easy traceability.

[0043] (2) Matching Recommendation Algorithm: Parts Database Construction: The system has a built-in parts database that stores information such as model specifications, compatible transmission types, manufacturer information, repair process level, historical price data, warranty period, and delivery cycle of various new and remanufactured parts. The database is synchronized with the product data of mainstream auto parts suppliers in real time, with an update frequency of no less than once every 12 hours to ensure data timeliness. Matching Filtering Logic: Based on the parts category, transmission model compatibility range, and remanufacturing requirement level of the old parts, a feature vector of old parts demand is constructed. ,in Parts category codes, To adapt to the transmission model code, Encode the remanufacturing demand level; simultaneously construct feature vectors for products in the parts database. ,in The corresponding accessory category code for the product. To adapt the product to the transmission model code, The product is coded according to its remanufacturing requirement level. A cosine similarity algorithm is used to calculate the fit. The formula is: , The value range is [0,1]. The closer to 1, the higher the fit. Recommended sorting rule: by fit. The top 10 products with the highest suitability are sorted from highest to lowest and selected as recommendations. Key information such as product type (new or remanufactured), manufacturer, historical price range, warranty period, and delivery time is also noted. The recommendation priority rules are as follows: remanufactured parts (meeting environmental and cost control requirements) are recommended first; when the remanufacturing requirement level is three (complex repair), corresponding new parts are also recommended for stores to choose from; in cases of equal suitability, products with a high degree of overlap between their historical price range and the remanufacturable value range of the old parts are prioritized.

[0044] (3) Integration of template and recommendation results: The standardized inquiry template is linked and integrated with the recommended parts list to generate a complete inquiry package, which includes three parts: digital inquiry template, parts recommendation list, and recommendation reasons. The recommendation reasons include suitability analysis, price matching analysis, and delivery cycle analysis, which are available for stores to view and push. After the inquiry package is generated, it is notified to store users through the APP. After the users confirm, it can be pushed to the parts suppliers.

[0045] Example: Following the valuation results of the Volkswagen DSG 7-speed dual-clutch transmission valve body obtained in the previous step, generate a quote template and recommended parts. First, fill in the content of each module in the template: Used Part Basic Information Module: Part category is dual-clutch valve body; transmission model compatibility is Volkswagen DSG 7-speed dual-clutch transmission; estimated service life of the used part is 5 years based on image analysis and brand identification; material is aluminum alloy. Damage Details Module: Damage type is crack; crack length is 4.5 mm; crack width is 0.5 mm; wear depth is 0.3 mm; no fatigue spalling, deformation, or oil passage blockage; damage area is described as the area near the oil passage. Value Range Module: Remanufacturable value range is 890-1044 yuan. Remanufacturing Demand Level Module: Value deduction rate. The value is in the range of 0.3-0.6, and the remanufacturing demand level is Level 2 (routine repair). Expected quotation requirements module: The default selection is a 7-day quotation validity period, cash on delivery, and SF Express logistics delivery. No modification is needed upon store confirmation.

[0046] The system invokes a parts matching and recommendation algorithm to construct a feature vector of old parts demand. (001 is the dual-clutch valve body code, 102 is the Volkswagen DSG 7-speed transmission code, and 02 is the secondary repair code). Similarity calculations were performed between these codes and the product feature vectors in the parts database. The 10 products with the highest compatibility were selected, of which 8 are remanufactured parts and 2 are new parts. The recommended list of remanufactured parts includes dual-clutch valve bodies from manufacturer A (compatibility 0.98, historical price range 850-1000 RMB, 1-year warranty, 3-day delivery time) and dual-clutch valve bodies from manufacturer B (compatibility 0.95, historical price range 880-1020 RMB, 1-year warranty, 4-day delivery time). The recommended list of new parts includes dual-clutch valve bodies from manufacturer C (compatibility 0.96, market price 1800 RMB, 2-year warranty, 2-day delivery time) and dual-clutch valve bodies from manufacturer D (compatibility 0.93, market price 1750 RMB, 2-year warranty, 3-day delivery time). The system generates a complete inquiry package, including a standardized inquiry template, a recommended parts list, and compatibility analysis. Once the store confirms the package, it can be sent to the parts supplier.

[0047] The unique technical approach of this solution lies in the standardized integration of the characteristic information, value range, and remanufacturing demand level of used parts to construct a structured inquiry template. This template clearly defines the complete fields and generation rules, and utilizes a cosine similarity algorithm based on multi-dimensional feature vectors to achieve precise matching and recommendation of parts. It also clarifies recommendation priorities and ranking rules, rather than relying on traditional free text inquiries or single-type part recommendations. Compared to existing technologies with fragmented information and ambiguous matching, this step ensures the completeness and standardization of inquiry information through standardized templates, reducing the invalid quote rate from parts suppliers by over 70%. Precise matching and recommendation further reduces invalid quotes from parts suppliers, improving the targeting and accuracy of quotes. Simultaneously, it provides stores with diverse options, shortens the communication cycle for inquiries and quotes, and improves transaction efficiency by over 60%.

[0048] (V) Parts Supplier Response Steps: The parts supplier response steps are a crucial link in the system's transaction loop. Their purpose is to push the generated inquiry packages to suitable parts suppliers, receive their quotes, and provide feedback to the store, achieving efficient interaction between the store and parts suppliers. In traditional parts supplier interactions, stores need to manually contact multiple suppliers, resulting in high communication costs, untimely information transmission, and a lack of unified reference standards for supplier quotes. This step, through parts supplier capability profiling, targeted push algorithms, standardized quote reception, and two-way negotiation functions, achieves precise reach of inquiry packages and standardized management of quote information. All functions are designed specifically for the needs of used gearbox parts transactions and possess strong process adaptability. Specific technical means are as follows: (1) Construction of Parts Supplier Qualification and Capability Profile: Data Collection: The system collects information such as business licenses, remanufacturing qualification certificates, business scope, historical transaction records, repair process capabilities, customer evaluations, delivery time data, and warranty performance of parts suppliers to establish a basic database of parts suppliers. Repair process capabilities are verified through repair cases uploaded by parts suppliers (images of old parts before repair + images after repair + inspection reports). Customer evaluations are submitted by store users after the transaction is completed. Delivery time data and warranty performance are automatically statistically analyzed through system transaction records. Capability Profile Generation: Hierarchical clustering algorithm is used to analyze five dimensions of parts suppliers: repair process capabilities, range of compatible products, reasonable pricing, delivery time, and credit rating to generate a parts supplier capability profile. The profile includes key indicators such as core repair types (e.g., valve body crack repair, gear wear repair), compatible gearbox brands (e.g., Volkswagen, Toyota), price range, average delivery cycle, and credit rating (1-5 stars, calculated based on customer evaluations and transaction records). The capability profile is updated quarterly to ensure its accuracy.

[0049] (2) Targeted Push Algorithm: Push Filtering Logic: Based on the parts supplier capability profile and the old parts requirements (parts category, remanufacturing requirement level, compatible transmission model) in the inquiry package, parts suppliers that meet all of the following conditions are selected: 1. The core repair type includes the damage type of the old part; 2. The compatible transmission brand includes the compatible brand of the old part; 3. The historical price range overlaps with the remanufacturable value range of the old part; 4. Reputation level ≥ 3 stars (out of 5 stars); 5. Average delivery cycle ≤ 7 days. Push Priority Ranking: The push is weighted according to three dimensions: the matching degree between the parts supplier and the old parts requirement (weight 0.5), the historical transaction rate (weight 0.3), and the delivery timeliness (weight 0.2). The weighted score is as follows: ,in To ensure a good match between parts suppliers and the demand for used parts, This represents the historical transaction rate of parts suppliers. This represents the average delivery cycle for parts suppliers. The top 10 parts suppliers with the highest matching degree are selected as the target audience for targeted push notifications. Push methods: Inquiry package links are sent to parts suppliers via the system's built-in message notification module through three methods: app push, SMS, and email, ensuring timely receipt by parts suppliers. Push records are automatically stored in the system, including push time, push method, and whether the message has been read, for easy traceability.

[0050] (3) Quotation Receipt and Management: Quotation Submission Function: Parts suppliers can click the inquiry package link to view the complete inquiry template and recommended parts list, and submit quotations based on the recommended parts or by selecting compatible parts independently. The quotation interface requires filling in information such as quotation amount, warranty period, delivery cycle, payment method, and logistics plan. It supports uploading additional documents such as product testing reports and repair process descriptions. The quotation amount must be within a reasonable range (50%-150% of the remanufacturing value of old parts), otherwise it cannot be submitted. Quotation Verification: The system verifies the quotations submitted by parts suppliers. The verification includes whether the quotation amount is a valid value, whether the warranty period is within a reasonable range (1-2 years for remanufactured parts, 2-3 years for new parts), and whether the delivery cycle is within a reasonable range (1-7 days). After the verification is passed, it is stored in the quotation database and a unique quotation number is generated; if the verification fails, the parts supplier is prompted to modify and resubmit. Quotation Feedback: The system pushes the quotation information of parts suppliers to the store APP in real time, and displays it in order of quotation amount from low to high. At the same time, it marks the parts supplier's credit rating, historical transaction rate, delivery cycle and other information to facilitate comparison and selection by the store. Stores can view price details and attached documents, and can save and annotate prices.

[0051] (4) Negotiation and Transaction Completion: Negotiation Function: Stores can select a quote to initiate negotiation, fill in the expected quote amount and the reason for negotiation (e.g., "The price is higher than the market average, and we expect it to be adjusted to XX yuan"). The system will push the negotiation information to the corresponding parts supplier, who can choose to accept, reject, or propose a new quote, forming a two-way negotiation process. Negotiation records are automatically stored in the system, including negotiation time, negotiation content, and responses from both parties. Transaction Confirmation: After the store confirms the final selected quote, the system generates an electronic transaction contract. The contract includes information on used parts, quote information, warranty terms, logistics terms, payment terms, and liability for breach of contract. The contract is confirmed using an electronic signature, which complies with the relevant requirements of the Electronic Signature Law and has legal effect. After both parties sign and confirm online, the transaction is completed. The system sends a transaction confirmation notification to both parties and generates a logistics order number, connecting with the logistics platform for logistics tracking.

[0052] Example: Following the Volkswagen DSG 7-speed dual-clutch valve body inquiry package generated in the previous step, the parts supplier response process is implemented. First, the parts supplier database is queried to build a supplier capability profile, and 10 suppliers meeting the following criteria are selected: core repair types include valve body crack repair, compatibility with Volkswagen brand transmissions, historical price range overlaps with 890-1044 yuan, reputation rating ≥ 3 stars, and average delivery cycle ≤ 7 days. These suppliers are then sorted according to matching degree (weight 0.5), historical transaction rate (weight 0.3), and delivery timeliness (weight 0.2) to determine the top 10 targeted suppliers. The system sends the inquiry package link to these 10 suppliers via APP push, SMS, and email.

[0053] After receiving the inquiry package, Parts Supplier A reviewed the inquiry template and recommended list for dual-clutch valve bodies, selected the compatible remanufactured part produced by themselves, and submitted a quote: 950 RMB, 1-year warranty, 3-day delivery period, cash on delivery, and free shipping via SF Express. They also uploaded a description of the laser cladding repair process and a test report for the remanufactured part. After the system verified the quote information, it was stored in the quote database and pushed to the store's app in real time. The store app displayed quotes from various suppliers sorted by price. Supplier A's quote of 950 RMB was at a mid-range level, with a 4.8-star reputation rating, a 92% historical transaction rate, and a 3-day delivery period. The store was relatively satisfied with the quote and initiated a negotiation, hoping for a price of 900 RMB, citing that "the average transaction price for similar used parts in the past 3 months was around 900 RMB." After receiving the negotiation information, Supplier A agreed to adjust the quote to 900 RMB. The store confirmed acceptance of the quote, the system generated an electronic transaction contract, both parties signed and confirmed online, the transaction was completed, the system generated a logistics tracking number and connected with SF Express to track the logistics status.

[0054] This technical solution's unique approach lies in its targeted push mechanism built upon parts supplier capability profiles. It clearly defines screening criteria and priority rules to ensure precise reach of inquiry packages. Simultaneously, it establishes a complete transaction interaction module encompassing quotation verification, sorting and display, two-way negotiation, and electronic contract signing, clearly defining operational rules and data requirements for each stage, rather than relying on traditional generalized push notifications or simple quotation reception functions. Compared to the problems of indiscriminate push notifications, chaotic quotation management, and cumbersome negotiation processes in existing technologies, this approach reduces invalid receipt by parts suppliers through precise targeted push notifications, increasing the quotation response rate by over 80%. Standardized quotation management and sorting and display facilitate quick comparison and selection by stores. Two-way negotiation and electronic contract signing promote reasonable transaction prices between both parties, reducing the transaction dispute rate by over 70%, significantly improving the efficiency and success rate of transaction matching, and reducing communication costs and transaction risks.

[0055] (vi) Model Iteration Steps: The purpose of the model iteration steps is to collect transaction data and feedback data during the actual operation of the system, and to dynamically optimize the MT-CNN model for image processing and the damage-remanufacturing value mapping model for value assessment, ensuring the recognition accuracy and valuation accuracy of the system during long-term operation. Traditional systems have fixed models after deployment, which cannot adapt to market changes and new damage types or old part models, leading to performance degradation after long-term use. This step achieves continuous self-evolution of the model through precise data collection, targeted deviation analysis, and scientific incremental training. All iteration processes are clear and repeatable, ensuring the effectiveness and stability of model optimization. Specific technical methods are as follows: (1) Data collection and storage: Data types: The collected feedback data includes three categories: image recognition deviation data, value assessment deviation data, and transaction feedback data. All data is authentic and traceable. Image recognition deviation data: After receiving the used parts, the parts supplier takes a picture of the actual product and uploads it to the system. The system compares the recognition results of the actual product image with the image uploaded by the store and calculates the deviation value of the damage quantification index, such as the crack length deviation. Wear depth deviation etc., among which The crack length is identified from the physical image. Wear depth for physical image recognition. Valuation deviation data: Collect information on the deviation between actual transaction prices and the valuation range, and calculate the deviation rate. ,in This is the actual transaction price. This represents the midpoint of the value assessment range. Transaction feedback data: Information such as transaction ratings (1-5 stars), repair quality feedback (acceptable / unacceptable), and delivery timeliness feedback (on time / delayed) from stores and parts suppliers is collected as a reference for model optimization. Data storage: Feedback data is stored using a distributed database (Hadoop), with data indexes established. Index fields include old part type, transaction time, and deviation rate to ensure efficient data querying and retrieval. Simultaneously, the data is encrypted using the AES-256 encryption algorithm to ensure data security and comply with relevant requirements of the Data Security Law.

[0056] (2) Deviation analysis and sample selection: Deviation analysis: Statistical analysis of image recognition deviation data is performed to calculate the deviation rate of various damage quantification indicators. ,in This is the deviation value of the quantitative indicator. This is a quantitative indicator value for object recognition. Preset deviation rate threshold. Samples with a deviation rate exceeding the threshold are selected, and the reasons for the deviation are analyzed, such as improper shooting angle, incorrect damage type identification, or unreasonable quantization algorithm parameters. The samples are then classified and labeled. Value assessment deviation analysis: The value assessment deviation data is analyzed, and a preset value assessment deviation rate threshold is established. Samples with deviation rates exceeding a certain threshold are selected, and the reasons for the deviations are analyzed, such as unreasonable weighting of value influencing factors, untimely updates of market adaptation factors, and inaccurate repair process cost data. The samples are then categorized and labeled. Sample labeling and expansion: The selected deviation samples are re-labeled, including the correct accessory category, damage type, damage quantification indicators, and value assessment results. These are then added to the model training dataset to expand the dataset's diversity and relevance. Each expansion includes at least 5% of the original training set.

[0057] (3) Incremental training and optimization of the model: MT-CNN model optimization: Incremental training is adopted, adding the expanded dataset to the original training set while keeping the original model parameters unchanged. Only the new samples are trained and updated, and the feature extraction weights of the model are adjusted. During training, the original Adam optimizer and multi-task loss function are used. The number of iterations is 1 / 5 of the original training number (10 epochs). The initial learning rate is 0.0001, which decays by 10% every 2 epochs to avoid overtraining and overfitting of the model. After training, the model performance is verified using a test set. The average deviation rate of various damage quantification indicators is required to be reduced to below 8%. Otherwise, retraining is required. Value mapping model optimization: Based on the value assessment deviation analysis results, the weight coefficients of the value influence factors are adjusted. Update the real-time data of market adaptation factors and the repair process cost database, and retrain the GBDT model with the same training parameters as the initial training. After training, use a test set to verify the model performance, requiring the value assessment bias rate to be reduced to below 8%; otherwise, retraining is required. Model deployment and verification: Deploy the optimized model to the system, set a model switching transition period (15 days), during which the new model and the old model run in parallel, comparing their recognition accuracy and valuation accuracy; after the transition period, verify the model performance using a new test dataset, ensuring that the recognition accuracy and valuation accuracy are significantly improved compared to before optimization. After successful verification, officially replace the old model and put it into use, and archive and store the old model.

[0058] Example: After the system has been running for 3 months, the model iteration process is initiated. First, feedback data is collected: a total of 500 transaction feedback data points are collected, including 300 data points of image recognition deviation, 500 data points of value assessment deviation, and 500 transaction feedback data points.

[0059] Analyzing the image recognition deviation data, the deviation rates of various damage quantification indicators were calculated. Twenty samples with crack length deviation rates exceeding 10% were selected. Analysis revealed that these samples were mostly internal cracks in oil passages, which could not be fully captured by the shooting angle, resulting in the recognized crack length being less than the actual length. These samples were re-labeled and added to the training dataset of the MT-CNN model. Incremental training was employed, adding the 20 new samples to the original training set, using the Adam optimizer and multi-task loss function, iterating for 10 epochs with an initial learning rate of 0.0001, decreasing by 10% every two epochs to optimize the MT-CNN model. After optimization, validation was performed on a new test set. The recognition deviation rate for internal cracks in oil passages decreased from 15% to 8%, and the average recognition deviation rate for overall damage quantification indicators decreased from 9% to 6%.

[0060] Analysis of valuation deviation data identified 30 samples with a deviation rate exceeding 15%. These samples were found to be pre-owned parts for a new type of transmission, where outdated average supply-demand ratio data for market adaptability factors led to undervaluation. The solution is to update the real-time supply-demand ratio data for market adaptability factors and adjust the weighting coefficients of value-influencing factors. The market fit factor weight was increased from 0.2 to 0.25, and the GBDT model was retrained. After optimization, the value assessment bias rate on the new test set decreased from 12% to 8%, and the accuracy of value assessment was significantly improved.

[0061] The two optimized models were deployed to the system, with a 15-day transition period during which the new and old models ran in parallel. Comparison revealed that the new model improved recognition accuracy by 8% and reduced value assessment bias by 5%. After the transition period, verification was performed using 100 new transaction data points. Overall recognition accuracy improved by 8%, value assessment bias decreased by 5%, and transaction satisfaction with stores and parts suppliers increased by 10%. The model iteration achieved the expected results and was officially put into use, replacing the old model.

[0062] The unique technical approach of this solution lies in constructing a closed-loop model iteration system based on transaction feedback data. It clearly defines the data collection scope, bias analysis thresholds, sample selection rules, and incremental training parameters. Instead of traditional full retraining or manual parameter adjustment, it employs incremental training to specifically optimize the MT-CNN and value mapping models. Compared to existing technologies where models remain unchanged or optimization costs are high, this step ensures targeted model optimization through precise bias analysis and sample selection. Incremental training reduces training time and resource costs, improving training efficiency by over 70%. Continuous iteration enables the system to adapt to new damage types, older part models, and market changes, maintaining high identification accuracy and valuation accuracy over the long term, thus extending the system's lifespan and core competitiveness.

[0063] It should be noted that the technical means of this system refer to publicly available technologies in the field of computer vision (such as Gaussian filtering, histogram equalization, MT-CNN network architecture, GBDT algorithm, etc.) and publicly available data standards in the field of auto parts trading. However, there are significant differences in specific applications and optimizations. All technical means are designed specifically for the needs of the scenario of used automotive transmission parts.

[0064] In the image acquisition stage, publicly available image quality verification algorithms were referenced. However, by associating and matching the type of old parts with the shooting parameters, clarifying the specifications of auxiliary tools and the requirements of shooting angles, a multi-dimensional verification mechanism was constructed. This solved the unique problem of capturing reflections, oil stains, and minor damage in the shooting of old gearbox parts. This is a scenario-specific optimization that general image acquisition technologies have not addressed.

[0065] In the image processing stage, the publicly available MT-CNN network architecture was referenced, but by designing a multi-task output layer specific to old gearbox parts, damage type recognition and component function weight classification were added, and a scene-specific damage quantification model was constructed. This achieved accurate recognition and quantification of damage types specific to old gearbox parts, which is different from the single-task output of general object recognition and fully discloses the technical means.

[0066] In the valuation process, the publicly available GBDT algorithm and multi-factor evaluation concept were referenced, but a four-dimensional value influencing factor system adapted to used transmission parts was constructed. The quantitative formulas and parameter meanings of each factor were clarified, and a non-linear value mapping model was established. This solved the valuation problem of strong coupling between the value of used transmission parts and multi-dimensional factors. It is different from the linear or single-factor valuation model of general used parts and has outstanding substantive features and significant progress.

[0067] In the process of generating inquiry templates and responding to parts suppliers, publicly available document generation technologies and recommendation algorithms were referenced. However, through standardized template design, construction of parts supplier capability profiles, targeted push mechanisms, and two-way negotiation functions, the entire process of trading used gearbox parts was standardized and accurately matched. This is different from the generalized inquiry and push model of general e-commerce platforms. The technical means are clear and can be implemented repeatedly.

[0068] In the model iteration phase, publicly available incremental training techniques were referenced. However, through the precise collection of transaction feedback data, targeted deviation analysis, and scenario-specific sample selection, scenario-specific optimization of the model was achieved. This ensures that the system can continuously adapt to the unique needs of used gearbox parts trading. Unlike the indiscriminate iteration of general AI systems, the innovation and practicality of the technical solution are outstanding.

[0069] In summary, the key advantage of this technical solution lies in its deep integration of computer vision technology with the used gearbox parts trading scenario. Addressing the unique challenges of damage identification, value assessment, and transaction matching in this scenario, it utilizes scenario-specific algorithm optimization, model building, and process design to create a complete and practical intelligent valuation and trading system. This significantly improves transaction efficiency and accuracy, lowers the professional threshold, and solves many pain points of traditional manual methods, demonstrating strong scenario exclusivity and practical value.

Claims

1. An image recognition and intelligent valuation system for used automotive transmission parts, characterized by: The system includes an image acquisition module, an image processing module, a quotation template generation module, a parts supplier response module, and a model iteration module. The image acquisition module acquires image data of used transmission parts. The image processing module uses a multi-task convolutional neural network to extract the part category, damage area, damage type, and wear quantification index of the used transmission parts. It also calculates the remanufacturable value range of the used transmission parts by constructing a damage-remanufacturing value mapping model. The mapping relationship of the damage-remanufacturing value mapping model is as follows: ， in This represents the midpoint of the remanufacturable value range. The basic market value for undamaged used transmission parts. For the first The weighting coefficients of class value influencing factors, ， For the first The quantitative values ​​of the four value influencing factors are: damage parameter factor, component function weight factor, repair process cost factor, and market adaptability factor. The inquiry template generation module generates a standardized inquiry template based on the output of the image processing module and recommends matching new or remanufactured parts. The parts supplier response module pushes standardized inquiry templates to compatible parts suppliers, receives quotation information submitted by parts suppliers, and supports two-way negotiation between stores and parts suppliers. The model iteration module collects transaction prices, damage parameter deviations from physical verification feedback, and repair effect data from actual transactions. It then applies these data to the feature extraction weights of the multi-task convolutional neural network and the weight coefficients in the damage-remanufacturing value mapping model. Make dynamic adjustments.

2. The image recognition and intelligent valuation system for used automotive transmission parts as described in claim 1, characterized in that: The image acquisition module includes an image acquisition unit and an acquisition auxiliary unit. The image acquisition unit is used to capture multi-view images of the front panoramic view, side panoramic view, and close-up of the suspected damaged area of ​​the old gearbox part through a mobile terminal, and is accompanied by a solid color background and scale reference piece during shooting. The acquisition auxiliary unit analyzes the sharpness, illumination uniformity, and integrity of the old part of the captured image in real time through a sharpness verification algorithm, an illumination uniformity analysis algorithm, and an old part integrity verification algorithm, and outputs corresponding adjustment prompt information.

3. The image recognition and intelligent valuation system for used automotive transmission parts as described in claim 2, characterized in that: The image processing module includes an image preprocessing submodule, a feature extraction submodule, and a value calculation submodule. The image preprocessing submodule is used to perform Gaussian filtering for noise reduction, adaptive threshold segmentation and morphological opening operation combined with reflective oil stain suppression, histogram equalization enhancement, and contour correction on the image data acquired by the image acquisition module. The feature extraction submodule outputs the accessory category, damage type, location information of the damaged area, and wear quantification index through a multi-task convolutional neural network. The value calculation submodule calculates the remanufacturable value range based on the output of the feature extraction submodule through a damage-remanufacturing value mapping model. When the repair cost is greater than 1.2 times the median value of the remanufacturable value range, the old part is determined to be unremanufacturable and the value range is output as 0.

4. The image recognition and intelligent valuation system for used automotive transmission parts as described in claim 3, characterized in that: The damage types output by the feature extraction submodule include cracks, wear, fatigue spalling, deformation, and oil passage blockage. The wear quantification indicators include crack length, crack width, wear depth, fatigue spalling area, and the percentage of oil passage blockage area.

5. The image recognition and intelligent valuation system for used automotive transmission parts as described in claim 4, characterized in that: The inquiry template generation module includes a template construction unit and a matching recommendation unit. The template construction unit is used to integrate the parts category, damage type, damage area, wear quantification index, and remanufacturable value range to generate a standardized inquiry template that includes parts specifications, remanufacturing requirement level, quotation validity period, and logistics requirements. The matching recommendation unit uses a cosine similarity algorithm to select and recommend the new or remanufactured parts with the highest compatibility from a preset parts database based on the parts category, remanufacturing requirement level, and compatible gearbox model. Remanufactured parts are recommended first, and corresponding new parts are recommended simultaneously when the remanufacturing requirement level is complex repair.

6. The image recognition and intelligent valuation system for used automotive transmission parts as described in claim 5, characterized in that: The parts supplier response module includes a template push unit, a quotation receiving unit, and a price negotiation unit; The template push unit pushes standardized inquiry templates to suitable parts suppliers based on their business scope, repair capabilities, and historical transaction data. The quotation receiving unit is used to receive and store quotation information such as quotation amount, warranty period and delivery cycle submitted by parts suppliers for recommended new or remanufactured parts; The negotiation unit allows stores to initiate negotiation requests to parts suppliers and submit their expected quotes, and to receive and display the parts suppliers' negotiation feedback to the stores.

7. The image recognition and intelligent valuation system for used automotive transmission parts as described in claim 6, characterized in that: The model iteration module includes a data collection unit, a deviation analysis unit, and a parameter adjustment unit. The data collection unit collects actual transaction prices, deviation values ​​of damage parameters between physical objects and image recognition, quality inspection data of repaired old parts, and transaction evaluation feedback data. The deviation analysis unit statistically analyzes the deviation rate of damage parameter recognition and the deviation rate of value assessment, and filters samples with deviation rates exceeding a preset threshold. The parameter adjustment unit, based on the filtered samples, uses incremental training to adjust the feature extraction weights of the multi-task convolutional neural network and the weight coefficients in the damage-remanufacturing value mapping model. Updates and adjustments will be made.

8. The image recognition and intelligent valuation system for used automotive transmission parts as described in claim 7, characterized in that: The parts database stores the model specifications, compatible transmission types, manufacturer information, repair process levels, and historical price data of various new and remanufactured parts. The matching and recommendation unit compares and filters the data in the parts database based on the parts category, remanufacturing requirement level, and compatible transmission model, and outputs the top 10 recommendations with the highest compatibility and the reasons for the recommendations.

9. The image recognition and intelligent valuation system for used automotive transmission parts as described in claim 1, characterized in that: The quantification method for the four types of value-influencing factors is as follows: damage parameter factor ,in , , , , The weighting coefficients for cracks, wear, fatigue spalling, oil passage blockage, and deformation are respectively, and their sum is 1. , , , , These are the normalized values ​​of the quantitative indicators corresponding to the damage; Component function weight factor Preset weighting coefficients for core or secondary functional components; Repair process cost factor , The cost of the repair process corresponding to the damage, The market price of this type of new part; market fit factor: ， For adjustment coefficients, For the supply and demand ratio of used parts, This represents the average supply-demand ratio for this type of used parts.

Images