Vehicle residual value prediction method and device, electronic equipment and storage medium

Abstract

The invention relates to a vehicle residual value prediction method and device, electronic equipment and a storage medium, and the method comprises the steps: processing vehicle non-image data of a vehicle, obtaining to-be-processed data features, inputting the to-be-processed data features into at least two basic models, obtaining at least two initial prediction results, and inputting the initial prediction results into an integrated model to obtain a basic residual value prediction result; inputting vehicle image data of a vehicle into the visual language large model to obtain an image correction factor feature vector; and splicing the basic residual value prediction result and the image correction factor feature vector as an input feature vector, inputting the input feature vector into a regression model, and processing to obtain a target residual value prediction result of the vehicle. According to the technical scheme, the prediction results of the multiple basic models are fused through the integrated model to obtain the basic residual value prediction result, the basic residual value prediction result and the image correction factor feature vector are spliced and then input into the regression model for final prediction, the target residual value prediction result is obtained, and the vehicle residual value prediction accuracy is improved.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of automobile evaluation, and particularly relates to a vehicle residual value prediction method and device, an electronic device and a storage medium. BACKGROUND

[0002] With the vigorous development of the automobile industry, the automobile population has reached an unprecedented size, and at the same time, the size of the used car market is also growing rapidly. Vehicle residual value estimation in the process of used car transaction is particularly important.

[0003] Traditional personal experience-based evaluation of used car residual value is highly subjective and lacks objectivity, rationality and scientificity. Alternatively, a used car residual value estimation formula is used for estimation, or a simple statistical learning model or machine learning algorithm is used for estimation through mathematical modeling. Compared with the subjective experience-based evaluation of the evaluator, the estimation is more scientific, rational and objective, but the performance of a single model is insufficient, which also leads to inaccurate vehicle residual value estimation. SUMMARY

[0004] To solve the above technical problems or at least partially solve the above technical problems, at least one embodiment of the present disclosure provides a vehicle residual value prediction method and device, an electronic device and a storage medium.

[0005] In a first aspect, the present disclosure provides a vehicle residual value prediction method, comprising: acquiring multi-source heterogeneous data of a vehicle; wherein the multi-source heterogeneous data comprises vehicle non-image data and vehicle image data; processing the vehicle non-image data to obtain to-be-processed data features, and inputting the to-be-processed data features into at least two basic models for processing to obtain at least two initial prediction results, and inputting the at least two initial prediction results into an integrated model for processing to obtain a basic residual value prediction result; inputting the vehicle image data into a visual language large model for processing to obtain an image correction factor feature vector; and inputting the basic residual value prediction result and the image correction factor feature vector as an input feature vector into a regression model for processing to obtain a target residual value prediction result of the vehicle.

[0006] In a second aspect, the present disclosure provides a vehicle residual value prediction device, comprising: a first acquisition module configured to acquire multi-source heterogeneous data of a vehicle; wherein the multi-source heterogeneous data comprises vehicle non-image data and vehicle image data; a first processing module configured to process the vehicle non-image data to obtain a to-be-processed data feature; a second processing module configured to input the to-be-processed data feature into at least two base models for processing to obtain at least two initial prediction results; a third processing module configured to input the at least two initial prediction results into an integrated model for processing to obtain a base residual value prediction result; a fourth processing module configured to input the vehicle image data into a visual language large model for processing to obtain an image correction factor feature vector; and a fifth processing module configured to splice the base residual value prediction result and the image correction factor feature vector as an input feature vector to input into a regression model for processing to obtain a target residual value prediction result of the vehicle.

[0007] In a third aspect, the present disclosure provides an electronic device, comprising: a processor and a memory; the processor is configured to execute any of the vehicle residual value prediction methods provided in the embodiments of the present disclosure by calling programs or instructions stored in the memory.

[0008] In a fourth aspect, the present disclosure provides a computer-readable storage medium, which stores programs or instructions, and the programs or instructions cause a computer to execute any of the vehicle residual value prediction methods provided in the embodiments of the present disclosure.

[0009] In a fifth aspect, the present disclosure provides a computer program product, which is used to execute any of the vehicle residual value prediction methods provided in the embodiments of the present disclosure.

[0010] The technical solutions provided in the embodiments of the present disclosure have at least the following advantages compared with the prior art: In the embodiment of the present disclosure, multi-source heterogeneous data of a vehicle is acquired; wherein the multi-source heterogeneous data comprises vehicle non-image data and vehicle image data; the vehicle non-image data is processed to obtain to-be-processed data features, and the to-be-processed data features are respectively input into at least two basic models for processing to obtain at least two initial prediction results, and the at least two initial prediction results are input into an integrated model for processing to obtain a basic residual value prediction result; the vehicle image data is input into a visual language large model for processing to obtain an image correction factor feature vector; the basic residual value prediction result and the image correction factor feature vector are spliced as an input feature vector to input into a regression model for processing to obtain a target residual value prediction result of the vehicle. By using the above technical solution, the integrated model is used to comprehensively obtain initial prediction results of multiple basic models to obtain a basic residual value prediction result, and the visual language large model is used to process vehicle image data to generate an image correction factor feature vector, and finally the basic residual value prediction result and the image correction factor feature vector are combined to input into a regression model for processing to obtain a target residual value prediction result of the vehicle, thereby further improving the vehicle residual value prediction accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0011] The accompanying drawings, which are incorporated herein and form a part of the specification, illustrate embodiments consistent with the present disclosure and, together with the description, further serve to explain the principles of the present disclosure.

[0012] In order to more clearly illustrate the technical solutions in the embodiments of the present disclosure or the prior art, brief introductions will be given to the drawings needed to be used in the embodiments or prior art descriptions. Obviously, for those skilled in the art, other drawings can also be obtained from these drawings without any creative effort.

[0013] Figure 1 A flowchart of a vehicle residual value prediction method provided by an exemplary embodiment of the present disclosure is shown in the figure. Figure 2 An example diagram of vehicle residual value prediction provided by an exemplary embodiment of the present disclosure is shown in the figure. Figure 3 A flowchart of a vehicle residual value prediction method provided by another exemplary embodiment of the present disclosure is shown in the figure. Figure 4 An example diagram of battery pack evaluation provided by an exemplary embodiment of the present disclosure is shown in the figure. Figure 5 An example diagram of battery field estimation provided by an exemplary embodiment of the present disclosure is shown in the figure. Figure 6 A structural diagram of a vehicle residual value prediction device provided by an embodiment of the present disclosure is shown in the figure. DETAILED DESCRIPTION

[0014] In order to enable a more clear understanding of the above-mentioned purposes, features and advantages of the present disclosure, the present disclosure will be further described in detail below in conjunction with the accompanying drawings and embodiments. It can be understood that the described embodiments are part of the embodiments of the present disclosure, rather than all the embodiments. The specific embodiments described herein are merely intended to explain the present disclosure, rather than limit the present disclosure. In the absence of conflicts, the embodiments of the present disclosure and the features in the embodiments can be combined with each other. Based on the described embodiments of the present disclosure, all other embodiments obtained by those of ordinary skill in the art belong to the scope of protection of the present disclosure.

[0015] In the following description, many specific details are set forth in order to provide a thorough understanding of the present disclosure, but the present disclosure can also be implemented in other different manners than those described herein; obviously, the embodiments in the description are only part of the embodiments of the present disclosure, rather than all the embodiments.

[0016] The existing vehicle residual value prediction method, whether through a used car residual value estimation formula or through mathematical modeling, uses a simple statistical learning model or machine learning algorithm to estimate the residual value of a used car. The single model prediction considers insufficient comprehensive factors, is not suitable for complex scene regression problems, and leads to inaccurate vehicle residual value estimation.

[0017] In view of the above problems, the present disclosure provides a vehicle residual value prediction method, which acquires multi-source heterogeneous data of a vehicle; wherein the multi-source heterogeneous data includes vehicle non-image data and vehicle image data; processes the vehicle non-image data to obtain to-be-processed data features, and inputs the to-be-processed data features into at least two basic models for processing to obtain at least two initial prediction results, and inputs the at least two initial prediction results into an integrated model for processing to obtain a basic residual value prediction result; inputs the vehicle image data into a visual language large model for processing to obtain an image correction factor feature vector; splices the basic residual value prediction result and the image correction factor feature vector as an input feature vector to input into a regression model for processing to obtain a target residual value prediction result of the vehicle. By using the above technical solution, the integrated model is used to comprehensively obtain the initial prediction results of multiple basic models to obtain a basic residual value prediction result, and the visual language large model is used to process the vehicle image data to generate an image correction factor feature vector. Finally, the basic residual value prediction result and the image correction factor feature vector are combined to input into a regression model for processing to obtain a target residual value prediction result of the vehicle, thereby further improving the vehicle residual value prediction accuracy.

[0018] The specific embodiments of the vehicle residual value prediction method, device, electronic equipment and storage medium of the present disclosure will be explained in detail below in conjunction with the accompanying drawings.

[0019] Figure 1A flowchart of a vehicle residual value prediction method provided for an exemplary embodiment of the present disclosure can be executed by a vehicle residual value prediction device provided by an embodiment of the present disclosure, which can be implemented by software and / or hardware and generally integrated in an electronic device.

[0020] As shown in Figure 1 The vehicle residual value prediction method can include the following steps: Step 101, acquiring multi-source heterogeneous data of a vehicle; wherein the multi-source heterogeneous data includes vehicle non-image data and vehicle image data.

[0021] The vehicle refers to a vehicle to be evaluated for residual value, which can be a fuel-powered, hybrid-powered, pure-electric-powered, fuel cell-powered vehicle, etc. Different multi-source heterogeneous data can be acquired according to different power types, such as battery data, charging data, etc. for new energy vehicles, i.e. vehicles powered by batteries, to further improve the accuracy of subsequent residual value evaluation.

[0022] In an embodiment of the present disclosure, multi-source heterogeneous data generally refers to data of different structures acquired from different sources, including vehicle non-image data and vehicle image data; wherein the vehicle non-image data is index value data (i.e. structured data), including vehicle static information (such as registration date, vehicle age, new car price, vehicle model, brand, vehicle registration location, etc.), vehicle networking basic data (such as mileage, battery temperature, voltage, current, and power during electricity use, etc.), vehicle condition information (such as vehicle historical collision information, three-electricity, chassis, structural damage information, maintenance records, etc.), driving behavior information (such as sudden acceleration, sudden deceleration, sudden turn, average vehicle speed, average motor revolutions, maximum vehicle speed, fatigue driving, late-night driving, 100 km sudden acceleration, 100 km sudden deceleration, etc.), etc. The vehicle non-image data can be selected according to the actual prediction scene to further meet the vehicle residual value prediction requirements.

[0023] The vehicle image data refers to image data of the vehicle taken by the user through the terminal, such as multi-angle images of the vehicle's appearance, such as front, back, side, and tires, for identifying scratches, dents, scratches, paint spraying, aging marks, accident deformation, and non-original modification, etc. The vehicle's interior image, such as seat, instrument panel, and central control components, for identifying aging degree, wear, water stains, etc. The vehicle image data can be selected according to the actual prediction scene to further meet the vehicle residual value prediction requirements.

[0024] In step 102, the vehicle non-image data is processed to obtain the to-be-processed data features, and the to-be-processed data features are input into at least two base models for processing to obtain at least two initial prediction results, and the at least two initial prediction results are input into an integrated model for processing to obtain a base residual value prediction result.

[0025] In the embodiments of the present disclosure, after obtaining the vehicle non-image data, the vehicle non-image data can be preprocessed, and the preprocessed vehicle non-image data can be feature extracted to obtain the to-be-processed data features.

[0026] The preprocessing of the vehicle non-image data includes data cleaning, data conversion, data description, etc. In the data cleaning stage, the abnormal value data is removed or replaced, the missing value is directly deleted if the missing rate is too high, and the missing rate is filled or replaced if the missing rate is acceptable; in the data conversion stage, the continuous type field with dimensional influence is normalized to make different indicators comparable, and the continuous type field operation also includes logarithmic transformation, scaling transformation, standardization, binning, etc., and the discrete type feature is encoded; in the data description stage, statistical quantities are calculated and data is visualized as needed.

[0027] Further, there are many ways to extract features from the preprocessed vehicle non-image data to obtain the to-be-processed data features. As an example, battery data is obtained from the vehicle non-image data, and the battery data is processed to obtain a battery group total score, charging data is obtained from the vehicle non-image data, and the charging data is processed to obtain a battery field, preset component information is obtained from the vehicle non-image data, and the preset component information is processed to obtain a comprehensive vehicle condition score, speed information is obtained from the vehicle non-image data, and the speed information is calculated to obtain a comprehensive driving behavior score, and one or more of the battery group total score, the battery field, the comprehensive vehicle condition score, and the comprehensive driving behavior score are used as the to-be-processed data features.

[0028] It should be noted that the specific calculation methods of the battery group total score, the battery field, the comprehensive vehicle condition score, and the comprehensive driving behavior score are described in the subsequent model training embodiments, and will not be described here in detail.

[0029] It should be noted that the pre-processed vehicle non-image data such as the license plate date, the vehicle age, the new car price, the vehicle model, the brand, the city, the region, and the mileage can be used as the basic data features, and the battery pack total score, the battery field, the comprehensive vehicle condition score, and the comprehensive driving behavior score obtained by processing the pre-processed vehicle non-image data can be used as the additional data features. Generally, the basic data features and the additional data features are both selected as the to-be-processed data features for subsequent processing.

[0030] In the embodiments of the present disclosure, at least two base models are obtained, and the base models can be selected and set as needed, such as using LightGBM (Light Gradient Boosting Machine, high-performance gradient boosting framework), XGBoost (EXtreme Gradient Boosting, optimized distributed gradient boosting library), and CatBoost (Categorical Boosting, gradient boosting decision tree machine learning library) as the base models. Further, the to-be-processed data features are sequentially input into each base model for processing to obtain multiple initial prediction results. That is, the more base models are set, the more initial prediction results can be obtained. The specific number of base models can be selected and set according to actual application needs. The initial prediction result refers to an initial prediction score value of a vehicle obtained by processing the to-be-processed data features by the base model.

[0031] In the embodiments of the present disclosure, the integrated model is used to fuse the multiple initial prediction results to obtain a base residual value prediction result. Specifically, a vehicle non-image data sample of a vehicle is obtained, the vehicle non-image data sample has a residual value label, the vehicle non-image data sample is processed to obtain a to-be-processed data feature sample, the to-be-processed data feature sample is input into at least two pre-constructed base models as a training data set to obtain multiple initial training results, the multiple initial training results are spliced into a stacked training feature vector and input into a pre-constructed integrated model for processing to obtain a base residual value training result, a first training loss value is calculated based on the base residual value training result and the residual value label, and when the first training loss value is greater than a preset first loss threshold, the model parameters of the pre-constructed integrated model are adjusted, and then the base residual value training result and the residual value label are re-obtained to calculate the first training loss value until the first training loss value is less than or equal to the preset first loss threshold to obtain the integrated model. One initial training result is used as one training feature vector, multiple training feature vectors are obtained, and the multiple training feature vectors are combined in a specific order to form a new training feature vector as a stacked training feature vector. The specific order can be an order of sorting according to the score values of the initial training results or an order of the time of obtaining the initial training results, and the specific order can be selected and set according to actual application scenarios.

[0032] In the embodiments of the present disclosure, in order to guarantee the model performance while avoiding the problem of excessive training times, a first loss threshold can be preset as a model stop training condition. When the first training loss value is less than or equal to the preset first loss threshold, it indicates that the current trained integrated model meets the processing task requirements, and the model parameters of the pre-constructed integrated model can be stopped from being adjusted to obtain the integrated model.

[0033] Specifically, the historical vehicle non-image data of a plurality of vehicles is pre-stored as vehicle non-image data samples, which are used for initial prediction by a plurality of base models and for training an integrated model, so that the trained integrated model can more accurately fuse the initial prediction results and obtain more accurate base residual value prediction results when processing in real time. The base residual value prediction result refers to a base prediction score value of a vehicle, which needs to be corrected according to vehicle image data to obtain a final residual value, i.e., a target residual value prediction result of the vehicle.

[0034] In the embodiments of the present disclosure, after obtaining the to-be-processed data features, the to-be-processed data features are input into at least two base models for processing to obtain at least two initial prediction results, and the at least two initial prediction results are input into an integrated model for processing to obtain a base residual value prediction result.

[0035] Step 103: inputting the vehicle image data into a visual language large model for processing to obtain an image correction factor feature vector.

[0036] The visual language large model includes an image encoder, a text encoder, a deep learning module, and a multi-task output module. It can be understood that the visual language large model is pre-trained. Specifically, vehicle image data samples and a preset prompt word are input into a pre-constructed visual language large model, the vehicle image data samples are encoded based on the image encoder to obtain image vector samples, the preset prompt word is encoded based on the text encoder to obtain text vector samples, the image vector and the text vector are spliced to obtain spliced vector samples, the spliced vector is processed based on the deep learning module to obtain output vector samples, the output vector samples are input into the multi-task output module to obtain a plurality of task result samples, each task loss value is determined based on the plurality of task result samples and image labels, and a total loss value is determined based on each task loss value and preset task weight information. When the total loss value is greater than a preset total loss threshold, the model parameters of the pre-constructed visual language large model are adjusted, and the total loss value is determined again until the total loss value is less than or equal to the preset total loss threshold, thereby obtaining the visual language large model. The preset total loss threshold can be used as a model stop training condition. When the total loss value is less than or equal to the total loss threshold, it indicates that the current trained visual language large model meets the processing task requirements, and the model parameters of the pre-constructed visual language large model can be stopped from being adjusted to obtain the visual language large model.

[0037] Specifically, after obtaining the vehicle image data, the vehicle image data is input into the visual language large model for processing to obtain an image correction factor feature vector. As an example, the vehicle image data and a preset prompt word are encoded based on an image encoder and a text encoder respectively to obtain an image vector and a text vector, the image vector and the text vector are spliced to obtain a spliced vector, the spliced vector is processed based on a deep learning module to obtain an output vector, and the output vector is processed based on a multi-task output module to obtain the image correction factor feature vector.

[0038] The image correction factor feature vector refers to an adjustment factor vector of a basic residual value prediction result, and includes one or more of a scratch score, a dent number, a modification judgment, an interior aging level, and water stain detection.

[0039] As an example, multi-angle vehicle images (such as vehicle head, vehicle tail, side, interior view angle, etc.) are obtained as vehicle image data, and an OpenCV module (a cross-platform computer vision and machine learning software library) is used to group and number the multi-angle vehicle images, standardize the direction and view angle so that the model can understand; for each image or image group, a natural language input preset prompt word that meets the task requirements is dynamically generated to guide the joint reasoning of the multi-modal model, such as the preset prompt word being “please identify and score the scratch degree, dent number, and whether there is a modification behavior” in the image; a pre-trained visual language large model is used to perform multi-task reasoning, output multiple task results, such as regression tasks (such as scratch score, dent number, etc.), binary classification tasks (such as modification judgment, water stain detection), and multi-classification tasks (such as interior aging level, etc.), and finally the structured output is merged, that is, the multi-angle vehicle image output is fused to generate an image correction factor training feature vector at the whole vehicle level, such as “scratch score is 0.62”, “dent number is 2”, “modification judgment” “modification judgment is moderate”, “interior aging level is false”, and “water stain detection is 0.58”, etc.

[0040] In step 104, the basic residual value prediction result and the image correction factor feature vector are spliced as an input feature vector to input into a regression model for processing to obtain a target residual value prediction result of the vehicle.

[0041] In the embodiments of the present disclosure, after obtaining the basic residual value prediction result and the image correction factor feature vector, the basic residual value prediction result and the image correction factor feature vector are spliced as an input feature vector to input into a pre-trained regression model for processing to obtain a target residual value prediction result of the vehicle.

[0042] In the embodiments of the present disclosure, the regression model is pre-trained. Specifically, the basis residual value training result and the image correction factor training feature vector are spliced as an input training feature vector to input the pre-constructed regression model, to obtain a target residual value training result. The second training loss value is calculated based on the target residual value training result and the residual value label. When the second training loss value is greater than a preset second loss threshold, the model parameters of the pre-constructed regression model are adjusted, and then the target residual value training result and the residual value label are re-acquired to calculate the second training loss value until the second training loss value is less than or equal to the preset second loss threshold to obtain the regression model. In order to ensure the model performance and avoid the problem of too many training times, the preset second loss threshold can be used as a model stop training condition. When the second training loss value is less than or equal to the preset second loss threshold, it indicates that the current trained regression model meets the processing task requirements, and the model parameters of the pre-constructed regression model can be stopped adjusting to obtain the regression model.

[0043] In the embodiments of the present disclosure, the regression model is used to predict the remaining practical value of the vehicle within the reasonable service life. It can be understood that the regression model can predict future values according to existing data. In the embodiments of the present disclosure, the regression model is established between the basis residual value training result and the image correction factor training feature vector and the target residual value training result. When a new vehicle needs to be predicted, the basis residual value prediction result and the image correction factor prediction feature vector of the new vehicle are input into the regression model, and the target residual value prediction result of the new vehicle can be obtained.

[0044] Specifically, the image correction factor feature vector generated by analyzing the multi-angle vehicle appearance and interior image is fused on the basis residual value prediction result output by the integrated model to correct the basis residual value prediction result, and the final residual value estimation value, i.e., the target residual value prediction result of the vehicle, is output.

[0045] The vehicle residual value prediction method of the embodiments of the present disclosure comprises: acquiring multi-source heterogeneous data of a vehicle; wherein the multi-source heterogeneous data comprises vehicle non-image data and vehicle image data; processing the vehicle non-image data to obtain to-be-processed data features, and inputting the to-be-processed data features into at least two basic models for processing to obtain at least two initial prediction results, and inputting the at least two initial prediction results into an integrated model for processing to obtain a basic residual value prediction result; inputting the vehicle image data into a visual language large model for processing to obtain an image correction factor feature vector; and inputting the basic residual value prediction result and the image correction factor feature vector as an input feature vector into a regression model for processing to obtain a target residual value prediction result of the vehicle. By using the above technical solution, the integrated model is used to comprehensively integrate the initial prediction results of multiple basic models to obtain a basic residual value prediction result, the visual language large model is used to process the vehicle image data to generate an image correction factor feature vector, and finally the basic residual value prediction result and the image correction factor feature vector are combined to input into a regression model for processing to obtain a target residual value prediction result of the vehicle, thereby further improving the vehicle residual value prediction accuracy.

[0046] As shown in the example, Figure 2 comprises the following steps: step 2.1 data acquisition, including numerical data (vehicle non-image data described above) and image data (vehicle image data described above); step 2.2 data preprocessing, including data feature processing and image processing; step 2.3 inputting the data feature processing result into multiple basic models, and fusing the processing results of all basic models through an integrated module to obtain a basic residual value prediction result, while performing step 2.4 to input the image processing result into a VLM (Visual Language Model) for multi-task scoring output (such as scratches, dents, modifications, aging, and water stains, etc.), to obtain an image correction factor feature vector, and finally performing step 2.5 to input the basic residual value prediction result and the image correction factor feature vector into a regression model to obtain a final residual value prediction result, i.e., a target residual value prediction result of the vehicle.

[0047] Specifically, first, in the data acquisition stage, multi-source heterogeneous data is collected through vehicle networking and user terminal uploading, including numerical features such as vehicle static information, vehicle condition information, and driving behavior information, and image features such as multi-angle appearance and interior photos; second, in the data preprocessing stage, the original numerical data is processed for missing values and outliers, feature encoding and standardization conversion to ensure data quality and consistency; in the feature engineering stage, additional features such as battery consistency index, battery health (SOH, State of Health), collision factor and driving behavior score are constructed, and feature selection and optimization are realized through correlation analysis and principal component analysis; then, in the basic model construction stage, based on the preprocessed numerical features, LightGBM is used as the main regression algorithm for training, combined with XGBoost and CatBoost models to form multi-model prediction results input into the ensemble model (Stacking) fusion stage; in this stage, the first layer generates prediction features through cross-validation by multiple basic models, and the second layer uses ElasticNet (ElasticNet) regression as a meta-learner to learn the combined weights output by each basic model to generate a basic residual value prediction result; subsequently, in the image processing module, based on the multi-modal large model LLaVA-1.5 architecture and through LoRA (Low-Rank Adaptation, Low-Rank Adapter) light optimization, multi-task learning is realized to output image correction factor feature vectors such as scratch score, number of dents, modification judgment, interior aging grade, and water stain detection, and all image inputs are encoded by image encoder and text encoder, combined with preset prompt words to guide structured label generation.

[0048] Finally, in the post-processing regression model stage, the basic residual value prediction result and the standardized image correction factor feature vector are spliced as input features, input into the ElasticNet regression model for correction to capture nonlinear and interaction effects, and output the target residual value prediction result of the vehicle. In this way, a residual value estimation scheme is realized for multi-source data fusion, feature engineering richness, model fusion optimization, and multi-modal perception post-processing, effectively improving the accuracy of new energy vehicle residual value prediction.

[0049] Based on the description of the foregoing embodiments, it is necessary to pre-train each model by using multi-source heterogeneous data of different vehicles as a data training set to ensure the accuracy of vehicle residual value prediction. The following describes an example of a vehicle residual value prediction method provided by an example embodiment of the present disclosure. Figure 3 Detailed description.

[0050] Figure 3 A flowchart of a vehicle residual value prediction method provided by another example embodiment of the present disclosure is shown. The method can be executed by a vehicle residual value prediction device provided by an example embodiment of the present disclosure, which can be implemented by software and / or hardware and generally integrated in an electronic device.

[0051] As Figure 3 shown, the vehicle residual value prediction method can include the following steps: Step 301, obtaining a plurality of source heterogeneous data samples of a vehicle; wherein the plurality of source heterogeneous data samples include vehicle non-image data samples and vehicle image data samples, and the plurality of source heterogeneous data samples have residual value labels.

[0052] Specifically, the plurality of source heterogeneous data samples are divided into vehicle non-image data samples and vehicle image data samples, i.e. numerical value data and image data, including vehicle body features, user behavior, environmental semantics, and perceived image content, etc.

[0053] Among them, the vehicle non-image data is an index value data (i.e. structured data), including vehicle static information (such as license plate date, vehicle age, new car price, vehicle model, brand, vehicle registration place, etc.), vehicle networking basic data (such as mileage, battery temperature, voltage, current, power during electricity use, etc.), vehicle condition information (such as vehicle historical collision information, three-electricity, chassis, structural parts, etc. appearance damage information, maintenance record, etc.), driving behavior information (such as sudden acceleration, sudden deceleration, sudden turn, average vehicle speed, average number of motor revolutions, maximum vehicle speed, fatigue driving, late-night driving, 100km sudden acceleration, 100km sudden deceleration, etc.) and the like.

[0054] Among them, the vehicle image data includes vehicle appearance images such as front, rear, side, and tire, etc. multi-angle, which are used to identify scratches, dents, scratches, paint spraying, aging traces, accident deformation, non-original modification, etc.; vehicle interior images such as seats, instrument panels, central controls, etc. components, which are used to identify aging degree, wear, water stains, etc.

[0055] Among them, the plurality of source heterogeneous data samples have residual value labels, and the plurality of source heterogeneous data samples correspond to the historical real residual value of the vehicle.

[0056] Step 302, processing the vehicle non-image data samples to obtain a to-be-processed data feature sample, and inputting the to-be-processed data feature sample into at least two pre-constructed basic models respectively to obtain an initial training result output by each pre-constructed basic model.

[0057] In the embodiments of the present disclosure, there are many ways to process the vehicle non-image data samples to obtain the to-be-processed data feature sample, which can be selected and set according to actual application.

[0058] As an example, the battery data sample is obtained from the vehicle non-image data sample, the plurality of preset battery index parameter values are calculated based on the battery data sample, the initial score value of each preset battery index is determined based on each preset battery index parameter value and the preset threshold value, the weight of each preset battery index is determined based on the preset fuzzy judgment matrix, and the total score of the battery pack is obtained as the to-be-processed data feature sample based on the calculation of each preset battery index initial score value and each preset battery index weight.

[0059] Specifically, for new energy vehicles, the battery occupies a key position, so the battery data sample can be obtained for analysis to determine the total score of the battery pack. More specifically, the preset battery index includes voltage consistency, temperature consistency, internal resistance consistency, capacity consistency, and power consistency. The five preset battery indexes not only affect each other, but also reflect the consistency of the battery pack. In order to improve the overall and accurate evaluation of the consistency of the battery pack, the consistency of the battery pack is evaluated based on the five indexes. The preset threshold value can be understood as a specific score range corresponding to different preset battery index parameter values, so as to determine the initial score value corresponding to the preset battery index parameter value according to which score range the preset battery index parameter value is in.

[0060] As shown in Figure 4 , first, the cloud charging data is obtained as the battery data sample, including time, current, voltage, and temperature, which is used to calculate a plurality of preset battery index parameter values, i.e., the parameter values of the five indexes. Based on the parameter values of the five indexes and the reasonable threshold value set, the unweighted score (such as the percentage system) of each preset battery index is calculated. Then, the combined weighting coefficient obtained by the FAHP (Fuzzy Analytic Hierarchy Process, Fuzzy Analytic Hierarchy Process) and the double full method is used to calculate the weighted score of each preset battery index. The fuzzy judgment matrix is set in advance, the weight is calculated using the fuzzy judgment matrix, and the normalized weight is obtained. The weight calculation formula is ; wherein n is the order of the matrix, such as 4 preset battery indexes, i.e., 4 orders, represents the i-th and j-th fuzzy judgment matrix data, and finally the obtained weight is multiplied by each unweighted score to obtain the weighted score); and finally, the total score for evaluating the consistency of the battery pack, i.e., the total score of the battery pack, is obtained.

[0061] As another example, the charging data sample is obtained from the vehicle non-image data sample, the voltage and residual capacity curve graph is obtained based on the preset spline interpolation method and the charging data sample, the filling capacity and the capacity change amount are determined according to the voltage and residual capacity curve graph, and the first battery field is determined as the to-be-processed data feature sample based on the filling capacity and the capacity change amount; or, when the voltage and residual capacity curve graph cannot be obtained based on the preset spline interpolation method and the charging data sample, the capacity difference and the capacity charging change amount are obtained based on the charging data sample, and the second battery field is determined as the to-be-processed data feature sample based on the capacity difference and the capacity charging change amount.

[0062] The battery field refers to a battery state of health (SOH) field. As shown in the example of FIG. 5, the method includes the following steps: 5.1 obtaining a charging data sample; 5.2 calculating a capacity difference; 5.3 determining whether a voltage and residual capacity curve graph can be fitted; if yes, 5.4 fitting the voltage and residual capacity curve graph; 5.5 determining a filling capacity and a capacity change amount according to the voltage and residual capacity curve graph; 5.6 calculating a first battery field as a to-be-processed data feature sample by the filling capacity and the capacity change amount, i.e., calculating an SOC difference between an SOC end and an SOC start, dividing the capacity difference by the SOC difference, and taking an average SOC difference of all the SOC differences as the first battery field; and if no, 5.7 directly obtaining a capacity difference and a capacity charging change amount to determine a second battery field as a to-be-processed data feature sample, i.e., calculating an SOC difference between an SOC end and an SOC start, dividing the capacity difference by the SOC difference, and taking the result as the second battery field. Figure 5

[0063] ​Specifically, first, the charging data samples are divided into data segments (cut according to the changes of battery voltage and current over time), then suitable charging data samples are screened out (according to the changes of battery voltage and current, the samples in which the characteristics of voltage and current change over time are in the charging state are screened out), then the delta Q (power difference) is calculated by ampere-hour integration, and on this basis, a judgment is made to see if the OCV (Open Circuit Voltage) and SOC (State Of Charge) relationship graph can be constructed (i.e. to see if the data points of OCV and SOC are sufficient to be plotted, with interpolation as the basis for judgment), if it can be constructed, then the voltage and remaining capacity curve fitting is performed by using the spline interpolation method (a smooth curve passing through a series of points is made by using a variable spline; the interpolation spline is composed of a plurality of polynomials, each of which is determined by two adjacent data points, so that any two adjacent polynomials and their derivatives (usually not including the ninth derivative) are continuous at the connection point), the parking charging process data in the process is calculated, if not, the battery health SOH field in the process is estimated according to the charging capacity and the SOC change, and the final calculation formula is as follows SOH (battery health) = ; wherein, is the charging capacity in the charging process, The uploaded data and the interpolation results are calculated respectively.

[0064] As another example, a preset component information sample is obtained from the vehicle non-image data sample, and classification is performed based on the preset component information sample to obtain each preset component level, each preset component score value is determined based on each preset component level and a preset scoring strategy, and a comprehensive vehicle condition score is calculated based on each preset component score value and each preset component weight, serving as a feature sample of the to-be-processed data.

[0065] Specifically, the chassis, structural parts, cover parts, three-electricity preset components, and maintenance records of the vehicle are classified, such as being divided into four levels of micro, light, medium, and heavy, scoring of each local component is performed, and then a comprehensive vehicle condition score is calculated by weighting, that is, the comprehensive vehicle condition score serves as a feature sample of the to-be-processed data. As another example, a speed information sample is obtained from the vehicle non-image data sample, and calculation is performed based on the speed information sample to determine driving data in a target time range, a plurality of driving parameters are determined based on the driving data, and a comprehensive driving behavior score is calculated based on each driving parameter and a preset parameter weight, serving as a feature sample of the to-be-processed data.

[0066] Specifically, the rapid acceleration, rapid deceleration and other fields are calculated through the change of speed in a unit of time, and the 100km rapid acceleration, 100km rapid deceleration and other fields are calculated in combination with the mileage, the rapid turn field is calculated through the speed change and lateral acceleration in a short time, the average vehicle speed, the maximum vehicle speed, the average motor revolution and other driving parameters are calculated through the aggregation of the speed, the revolution information and other driving data in a certain time range, the fatigue driving, the late-night driving and other driving parameters are calculated through the driving data in the time dimension, and finally the scores of the various driving parameters are weighted and averaged according to their weights in the total score to obtain a comprehensive driving behavior score result; the determination of the weights can be selected and set according to actual application needs.

[0067] In the embodiments of the present disclosure, additional data features such as battery health and battery consistency are derived for the difference of battery properties, additional data features such as crash score are derived for the past use condition of the vehicle, additional data features such as driving behavior score are derived for the driving behavior of the vehicle, and the model training accuracy is further improved, thereby improving the subsequent prediction accuracy.

[0068] In step 303, the vehicle image data sample and the preset prompt word are input into the pre-constructed visual language large model, the vehicle image data sample is encoded based on an image encoder to obtain an image vector sample, the preset prompt word is encoded based on a text encoder to obtain a text vector sample, the image vector and the text vector are spliced to obtain a spliced vector sample, and the spliced vector is processed based on a deep learning module to obtain an output vector sample.

[0069] In step 304, the output vector sample is input into a multi-task output module to obtain a plurality of task result samples, each task loss value is determined based on the plurality of task result samples and the image label, and a total loss value is determined based on each task loss value and preset task weight information. When the total loss value is greater than a preset total loss threshold, the model parameters of the pre-constructed visual language large model are adjusted, and the total loss value is recalculated until the total loss value is less than or equal to the preset total loss threshold, and the visual language large model is obtained.

[0070] Specifically, the multi-angle appearance and interior image of the vehicle is intelligently analyzed and scored to extract key visual factors affecting the vehicle valuation. The pre-constructed visual language large model can use a multi-modal large model LLaVA-1.5 (Large Language and Vision Assistant, an end-to-end large multi-modal model) as a basic architecture, and perform lightweight adaptation through LoRA (Low-Rank Adaptation, low-rank adapter optimization) fine-tuning technology to realize structured output of multiple tasks, including scratch score, number of dents, modification judgment, interior aging level, water stain detection and the like.

[0071] Specifically, the vehicle image data sample is a plurality of image samples (multi-angle), including vehicle appearance view angles such as front, rear, left, right, headlights, tires, etc.; vehicle interior view angles such as seats, center control, instrument panel, doors, etc.; the model can use LLaVA-1.5 (based on CLIP image encoder + LLaMA language model), and use LoRA mechanism to inject learnable low-rank parameters into the Attention (attention mechanism) and Feedforward (feedforward network) sublayers of the Transformer (deep learning network), freeze the backbone model, and only optimize the LoRA insertion layer and the multi-task output head, to improve training efficiency and reduce resource requirements; use multi-task target joint optimization of regression plus classification, such as task name is scratch score, task type is regression, example output is 0 to 1 value, task name is dent number, task type is regression, example output is integer, task name is modification judgment, task type is classification, example output is yes or no, task name is interior aging level, task type is multi-classification, example output is new, almost new, slight, moderate and severe, etc., task name is interior aging level, task type is multi-classification, example output is almost new, slight, moderate and severe, etc., and task name is water stain detection, task type is classification, example output is yes or no.

[0072] Specifically, the basic model LLaVA-1.5 plus Prompt (preset prompt) is used for prediction, and the structured score label (such as scratch score = 0.8) is preliminarily generated, and these initial label values are used as reference suggestion values for manual labeling; wherein, at least two labels are generated for each image, at least two labeling versions are generated, and the difference is automatically checked, if the difference value exceeds the set threshold (such as the difference value of dent number > 1), further audit the correctness, improve the accuracy of image labeling, and thus improve the accuracy of subsequent model training.

[0073] Specifically, after the vehicle image data sample is labeled, the corresponding preset prompt is dynamically generated for each image, such as "please judge whether there are scratches, dents or modification behaviors in the vehicle image, and give the scratch score (0~1), dent number and whether there is modification respectively"; so that each image can obtain the task processing result, and the loss value is calculated combined with the labeled label, and the weighted sum is performed according to the preset task weight information to obtain the total loss value; wherein, different task weights are set differently, such as the scratch score is weighted average based on the view angle weight, the dent number is the total dent number of the whole vehicle image, the modification judgment is the maximum value, the interior aging level is the maximum aging level, and the water stain detection is the maximum value, etc.

[0074] Specifically, the image size is unified to a preset size such as 224*224, each piece of data contains an image path plus a preset prompt plus a multi-task structured label; the label value is uniformly mapped as: a regression field (32-bit floating point number) and a classification field (one-hot encoding); based on an open source large model LLaVA-1.5, including an image encoder, a language encoder, a LoRA parameter injection layer, a fusion module and multi-task output heads (multi-task heads) for completing joint prediction of regression and classification tasks; wherein the image encoder is, for example, CLIP ViT-L / 14 (Visual Transformer, Large model, 14*14 patch size), which is used to divide the input image into 14*14 image blocks (patches) and send them into a Transformer for encoding, outputting an image representation vector sequence, and the output structure is CLS Token, that is, a full-image semantic summary vector; Patch Tokens, a local visual representation corresponding to each image block; the final output is: wherein N=196 represents the number of patches, and D=768 represents the dimension of each Token.

[0075] wherein the language (text) encoder is, for example, LLAMA (an open source large language model developed by Meta), which is used to convert the input natural language prompt into a Token sequence and embed it into a language feature vector; wherein the language input format supports templated question guidance (such as "please judge whether there is a scratch or a dent…"); the Token embedding and fusion method is, for example, using LLaMA Tokenizer to encode natural language; the image Token and the text Token are spliced to form a unified input sequence: wherein M is the number of text Tokens; the insertion position of the LoRA parameter injection layer is the Attention and Feedforward module in the Transformer, which can be replaced by a structure such as the original linear layer replaced by ; wherein, ; ​The scaling factor controls the output amplitude of LoRA; all original parameters of the backbone model are frozen (including ViT and LLAMA); only the LoRA parameters A, B and the multi-task output head are trained; finally, the multi-task output head sets different output decoding structures according to the task type, such as the task type is regression, the output layer structure is a linear layer (output dimension = 1) and the loss function is mean square error (MSE, Mean-Square Error), such as the task type is binary classification, the output layer structure is a Sigmoid Layer (output is probability) and the loss function is binary cross entropy loss (BCE, Binary Cross Entropy), and such as the task type is multi-classification, the output layer structure is a Softmax Layer (output is a class probability distribution) and the loss function is cross entropy loss (BCE, Binary Cross Entropy). Wherein, each task corresponds to an independent decoder, and the prediction results of all tasks share the front-layer image-text features, and the output end adopts a parallel structure. Wherein, the loss function design is usually for multi-task output, and a weighted loss combination is adopted, and the weight of each task can be adjusted through a dynamic balancing strategy to avoid interference of the secondary task to the primary task loss.

[0076] In the embodiments of the present disclosure, a lightweight fine-tuning scheme based on an open-source multi-modal model (LLaVA-1.5) and LoRA (low-rank adapter) is designed, and multi-task structured scoring of vehicle appearance and interior image is realized, including scratch score, dent number, modification judgment, interior aging grade, water stain detection, etc.

[0077] In step 305, the initial training results of each pre-constructed base model are spliced into a stacked training feature vector, the stacked feature vector is input into the pre-constructed integrated model for processing to obtain a base residual value training result, a first training loss value is calculated based on the base residual value training result and the residual value label, and when the first training loss value is greater than a preset first loss threshold, the model parameters of the pre-constructed integrated model are adjusted and the first training loss value is recalculated, until the first training loss value is less than or equal to the preset first loss threshold, and the integrated model is obtained.

[0078] In step 306, the data feature samples to be processed are input into at least two base models to obtain initial training results of each base model, the initial training results of each base model are spliced into a stacked training feature vector, and the stacked training feature vector is input into the integrated model for processing to obtain a base residual value training result.

[0079] In step 307, the vehicle image data sample is input into the visual language large model to obtain an image correction factor training feature vector, the base residual value training result and the image correction factor training feature vector are spliced as an input training feature vector input into the pre-constructed regression model to obtain a target residual value training result.

[0080] Step 308, based on the target residual value training result and the residual value label, a second training loss value is calculated, and when the second training loss value is greater than a preset second loss threshold, the model parameters of the pre-constructed regression model are adjusted, and then the second training loss value is recalculated until the second training loss value is less than or equal to the preset second loss threshold, and the regression model is obtained.

[0081] Specifically, the processed data feature samples to be processed are taken as the training data set, the residual value in the historical data is taken as the label, and the remaining fields are taken as the basic data features, including the date of taking the card, the vehicle age, the new car sales price, the vehicle model, the brand, the city, the region, the mileage, and additional data features such as the battery factor (battery health, battery group total score), the collision factor (comprehensive vehicle condition score), and the driving behavior factor (comprehensive driving behavior score).

[0082] For example, taking LightGBM as the basic model, the basic model is initialized and related parameters such as the learning rate (learning_rate) are set, which is used to balance the model complexity and fitting ability, and excessive learning rate will lead to overfitting, and too small learning rate will lead to underfitting; for example, the maximum depth (max_depth) is used to limit the depth of the decision tree to prevent overfitting; for example, the number of leaf nodes (num_leaves) is used to control the growth size of the decision tree, which affects the training speed and prediction performance, etc. The system default values are used first, the sklearn package is called, the training data set is trained, and the mean square error is taken as the evaluation index to measure the prediction performance of the model.

[0083] In addition, the best parameter configuration of the model is determined through grid search and cross-validation (the grid search is instantiated, multiple cross-validations are performed, the XGBoost model and the CatBoost model are constructed according to the same process, and multiple initial training results are obtained.

[0084] Specifically, the first layer of the model is composed of multiple basic models, and the input is the original training data set; the second layer of the model is trained by taking the output of the first layer of the basic model as the training data set.

[0085] Wherein, for the base model, the entire data set of the to-be-processed data feature sample is split into two groups of training data sets such as training data set A and training data set B according to a certain proportion such as a proportion of 8 to 2, and the training data set A is divided into five parts, namely train1, train2, train3, train4 and train5; the base model is selected, such as selecting LightGBM, XGBoost and CatBoost; such as the XGBoost model part, train1, train2, train3, train4 and train5 are used in turn as the validation set, and the remaining 4 parts are used as the training data set for model training, and training is performed on the training data set B, so that five initial training results of the XGBoost model on the training data set A and one initial training result B1 on the training data set B are obtained, and then the five initial training results are vertically overlapped and combined to obtain A1, and the LightGBM and CatBoost model parts are the same, LightGBM obtains A2 and B2, and CatBoost obtains A3 and B3.

[0086] Specifically, after the base model is trained, the initial training results of the three base models on the training data set are taken as three "features" (A1, A2, A3) respectively, and then an ElasticNet (elastic net regression) model is trained and established as a pre-constructed integrated model; more specifically, the cross-validation results output by the three groups of base models are spliced into stacked training feature vectors, and the real residual value is taken as the label to fit the ElasticNet; the predictions (B1, B2, B3) output by the base model on the training data set B are used as input, and the final predicted residual value is output by the ElasticNet, that is, the trained ElasticNet (elastic net regression) model is used to predict the "feature" values (B1, B2, B3) on the training data set B of the three base models, and the final predicted residual value is output as the base residual value training result.

[0087] In the embodiments of the present disclosure, the characteristics and differences of multiple base models are comprehensively considered by using an integrated fusion method, the influence of each factor of the vehicle on the vehicle damage is considered, and the overall loss estimation of the vehicle has better prediction accuracy.

[0088] Further, on the basis of the base residual value training result output by the integrated model, the image correction factor feature vector generated by the multi-angle vehicle appearance and interior image analysis is fused to correct the base residual value training result, and the final residual value estimate value, i.e., the target residual value training result, is output.

[0089] Specifically, the correction factor field includes, for example, a scratch score, a dent number, a modification judgment, an interior aging grade, and a water stain detection, etc. In order to unify the scale and processing logic, all the correction factor fields are numerically processed and standardized to form an image correction factor training feature vector for combining with the basic residual value training result for post-processing.

[0090] Specifically, the image correction factor training feature vector is spliced with the basic residual value training result to obtain an input training feature vector inputting a pre-constructed regression model. The pre-constructed regression model learns a weighted relationship between a main model output and the image correction factor training feature vector in a feature space, and captures non-linear and interaction effects. ElasticNet regression is used, and the formula is Res_final=β0+β1*Res_base(basic residual value training result)+Σβi*V_corr[i](image correction factor training feature vector); wherein βi is a preset weight.

[0091] Specifically, the training data of the pre-constructed regression model is X=[Res_base_train, V_corr_train], y=actual historical transaction target residual value, and the mean square error between the predicted residual value and the true residual value is minimized. Loss=MSE(y, Res_final) through cross-validation, the optimal model parameters α, L1_ratio, etc. are obtained; the integrated model output basic residual value training result and the image multi-task score result, i.e. the image correction factor training feature vector V_corr, are received, spliced into a standardized input vector, input into a post-processing regression model, and the final prediction Res_final, i.e. the target residual value training result, is generated.

[0092] It should be noted that in addition to the LightGBM, XGBoost and CatBoost described in the above embodiments, deep learning models, traditional machine learning regression models such as decision tree regression, random forest regression, and support vector machine regression, etc. can also be used.

[0093] In the embodiments of the present disclosure, a post-processing regression model is designed, the input is the output of the basic model and the image multi-task score feature vector, the ElasticNet regression is used to learn the weighted relationship between the residual value and the image factor in the feature space, the quantitative correction effect of the real use state of the vehicle on the residual value is captured, the individualized post-adjustment of the integrated model prediction is realized, and the prediction result is further improved.

[0094] In summary, the embodiments of the present disclosure introduce a LightGBM and other basic models more suitable for vehicle scenarios, which have stronger learning performance, higher accuracy for vehicle residual value estimation, stronger generalization ability, and greater expansion for application. In addition, additional data features that are more reasonable and more in line with actual vehicle conditions are introduced, such as the battery health and battery consistency scores that are more important for new energy vehicles. For new energy vehicles, the related attributes of the vehicle battery occupy completely different roles and positions, and their good or bad greatly affect the value of the vehicle. The health and consistency of the vehicle battery are scored, and then repeated training is performed with appropriate weights. The use of these new features in model estimation can more accurately estimate the residual value of the vehicle. In addition, data modeling and feature weight adjustment can be performed for different vehicles to ensure that the estimation method adapts to various vehicle models and power types, fully considering the characteristics of each vehicle model. The model fusion considers more comprehensive factors, uses an integrated fusion method to consider the characteristics and differences of multiple basic models, considers the influence of each factor on the vehicle damage, and has better prediction accuracy for the overall loss of the vehicle. Finally, multi-perspective vehicle images are introduced, and a large model is used to generate structured visual factors through multi-task reasoning to achieve fine-grained wear and tear / retrofit scoring. A post-processing regression model is proposed to combine numerical prediction results and image score vectors to achieve feature level fusion correction and improve vehicle residual value prediction accuracy.

[0095] To implement the above embodiments, the present disclosure further provides a vehicle residual value prediction device, which can be implemented by software and / or hardware.

[0096] Figure 6 The structure diagram of the vehicle residual value prediction device provided by an embodiment of the present disclosure is shown in Figure 3 As shown in the figure, the vehicle residual value prediction device 60 can include: A first acquisition module 610 is configured to acquire multi-source heterogeneous data of a vehicle; wherein the multi-source heterogeneous data includes vehicle non-image data and vehicle image data; A first processing module 620 is configured to process the vehicle non-image data to obtain data features to be processed; A second processing module 630 is configured to input the data features to be processed into at least two basic models for processing to obtain at least two initial prediction results; A third processing module 640 is configured to input the at least two initial prediction results into an integrated model for processing to obtain a basic residual value prediction result; A fourth processing module 650 is configured to input the vehicle image data into a visual language large model for processing to obtain an image correction factor feature vector; The fifth processing module 660 is configured to splice the base residual value prediction result and the image correction factor feature vector as an input feature vector to input into a regression model for processing to obtain a target residual value prediction result of the vehicle.

[0097] Optionally, the first processing module 620 is specifically configured to: obtain battery data from the vehicle non-image data, and process based on the battery data to obtain a battery group total score; obtain charging data from the vehicle non-image data, and process based on the charging data to obtain a battery field; obtain preset component information from the vehicle non-image data, and process based on the preset component information to obtain a comprehensive vehicle condition score; obtain speed information from the vehicle non-image data, and calculate based on the speed information to obtain a comprehensive driving behavior score; and use one or more of the battery group total score, the battery field, the comprehensive vehicle condition score, and the comprehensive driving behavior score as the to-be-processed data feature.

[0098] Optionally, the fourth processing module 650 is specifically configured to: encode the vehicle image data and a preset prompt word based on an image encoder and a text encoder of the visual language large model to obtain an image vector and a text vector; splice the image vector and the text vector to obtain a spliced vector, and process the spliced vector based on a deep learning module of the visual language large model to obtain an output vector; and process the output vector based on a multi-task output module of the visual language large model to obtain the image correction factor feature vector.

[0099] Optionally, the device further includes: a second acquisition module configured to acquire a plurality of source heterogeneous data samples of a vehicle; wherein the plurality of source heterogeneous data samples include vehicle non-image data samples, and the plurality of source heterogeneous data samples have residual value labels; a sixth processing module configured to process the vehicle non-image data samples to obtain to-be-processed data feature samples; a seventh processing module configured to input the to-be-processed data features into the at least two pre-constructed base models respectively to obtain initial training results output by each of the pre-constructed base models; and an eighth processing module configured to splice the initial training results output by each of the pre-constructed base models into a stacked training feature vector, input the stacked feature vector into a pre-constructed integrated model for processing to obtain a base residual value training result, calculate a first training loss value based on the base residual value training result and the residual value labels, and recalculate the first training loss value after adjusting model parameters of the pre-constructed integrated model when the first training loss value is greater than a preset first loss threshold value, until the first training loss value is less than or equal to the preset first loss threshold value, and the integrated model is obtained.

[0100] Optionally, the device further comprises: the device further comprises: a third acquisition module, configured to acquire a multi-source heterogeneous data sample of the vehicle; wherein the multi-source heterogeneous data sample comprises a vehicle non-image data sample and a vehicle image data sample, and the multi-source heterogeneous data sample has a residual value label; a ninth processing module, configured to input the vehicle non-image data sample corresponding to a to-be-processed data feature sample into the at least two base models respectively, to obtain an initial training result output by each base model; a tenth processing module, configured to splice the initial training result output by each base model into a stacked training feature vector, and input the stacked training feature vector into an integrated model for processing, to obtain a base residual value training result; an eleventh processing module, configured to input the vehicle image data sample into the visual language large model, to obtain an image correction factor training feature vector; a twelfth processing module, configured to splice the base residual value training result and the image correction factor training feature vector as an input training feature vector to input into a pre-constructed regression model, to obtain a target residual value training result; and a thirteenth processing module, configured to calculate a second training loss value based on the target residual value training result and the residual value label, and when the second training loss value is greater than a preset second loss threshold, adjust model parameters of the pre-constructed regression model and then recalculate the second training loss value, until the second training loss value is less than or equal to the preset second loss threshold, to obtain the regression model.

[0101] Optionally, the sixth processing module is specifically configured to: acquire a battery data sample from the vehicle non-image data sample; calculate a plurality of preset battery index parameter values based on the battery data sample; determine an initial score value of each preset battery index based on each preset battery index parameter value and a preset threshold; determine a weight of each preset battery index based on a preset fuzzy judgment matrix, and calculate a total score of a battery pack based on the initial score value of each preset battery index and the weight of each preset battery index, as the to-be-processed data feature sample.

[0102] Optionally, the sixth processing module is specifically configured to: acquire a charging data sample from the vehicle non-image data sample; when a voltage and a remaining power curve graph are acquired based on a preset spline interpolation method and the charging data sample, determine a charging capacity and a power change amount according to the voltage and the remaining power curve graph, and determine a first battery field based on the charging capacity and the power change amount as the to-be-processed data feature sample; or, when the voltage and the remaining power curve graph cannot be acquired based on the preset spline interpolation method and the charging data sample, calculate a power difference and a power charging change amount based on the charging data sample, and determine a second battery field based on the power difference and the power charging change amount as the to-be-processed data feature sample.

[0103] Optionally, the sixth processing module is specifically configured to: obtain a preset component information sample from the vehicle non-image data sample, and perform classification based on the preset component information sample to obtain each preset component level; determine each preset component score value based on the each preset component level and a preset score strategy; and perform calculation based on the each preset component score value and each preset component weight to obtain a comprehensive vehicle condition score as the to-be-processed data feature sample.

[0104] Optionally, the sixth processing module is specifically configured to: obtain a speed information sample from the vehicle non-image data sample, and perform calculation based on the speed information sample to determine driving data in a target time range; determine a plurality of driving parameters based on the driving data, and perform calculation based on each driving parameter and a preset parameter weight to obtain a comprehensive driving behavior score as the to-be-processed data feature sample.

[0105] Optionally, the vehicle image data sample is specifically an image label, and the pre-constructed visual language large model comprises an image encoder, a text encoder, a deep learning module and a multi-task output module. The device further comprises a visual language large model training module configured to input the vehicle image data sample and a preset prompt word into the pre-constructed visual language large model, encode the vehicle image data sample based on the image encoder to obtain an image vector sample, encode the preset prompt word based on the text encoder to obtain a text vector sample, splice the image vector and the text vector to obtain a spliced vector sample, process the spliced vector based on the deep learning module to obtain an output vector sample, input the output vector sample into the multi-task output module to obtain a plurality of task result samples, determine each task loss value based on the plurality of task result samples and the image label, and determine a total loss value based on each task loss value and preset task weight information. When the total loss value is greater than a preset total loss threshold value, the model parameters of the pre-constructed visual language large model are adjusted, and the total loss value is recalculated until the total loss value is less than or equal to the preset total loss threshold value, thereby obtaining the pre-trained visual language large model.

[0106] The vehicle residual value prediction device provided in the embodiments of the present disclosure can execute the vehicle residual value prediction method provided in the embodiments of the present disclosure, and has the function modules and beneficial effects corresponding to the execution method. The contents not described in detail in the device embodiments of the present disclosure can be referred to the description in any method embodiment of the present disclosure.

[0107] The embodiments of the present disclosure further provide an electronic device including a processor and a memory. The processor is configured to execute the steps of the vehicle residual value prediction method according to any of the preceding embodiments by invoking programs or instructions stored in the memory. To avoid repetition, the description is omitted here.

[0108] The embodiments of the present disclosure further provide a computer readable storage medium, which is non-transitory, and stores programs or instructions, which make a computer execute steps of each embodiment of the vehicle residual value prediction method. For the purpose of avoiding repetitive description, no further description is given here.

[0109] The embodiments of the present disclosure further provide a computer program product, which is used to execute steps of each embodiment of the vehicle residual value prediction method.

[0110] It should be noted that, in this document, the terms such as "first" and "second" are merely used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply that there is any such actual relationship or order between these entities or operations. Moreover, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or device. Without more limitations, the element defined by the statement "including a" does not exclude the presence of other identical elements in the process, method, article or device including the element.

[0111] The above description is merely specific embodiments of the present disclosure, enabling those skilled in the art to understand or implement the present disclosure. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present disclosure. Therefore, the present disclosure will not be limited to these embodiments described herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method of predicting a vehicle residual value, characterized by, The method comprises the following steps: acquiring multi-source heterogeneous data of a vehicle; wherein the multi-source heterogeneous data comprises vehicle non-image data and vehicle image data; processing the vehicle non-image data to obtain to-be-processed data features, and inputting the to-be-processed data features into at least two basic models for processing to obtain at least two initial prediction results, and inputting the at least two initial prediction results into an integrated model for processing to obtain a basic residual value prediction result; inputting the vehicle image data into a visual language large model for processing to obtain an image correction factor feature vector; inputting the basic residual value prediction result and the image correction factor feature vector as an input feature vector into a regression model for processing to obtain a target residual value prediction result of the vehicle.

2. The method of claim 1, wherein, The processing of the vehicle non-image data to obtain to-be-processed data features comprises the following steps: acquiring battery data from the vehicle non-image data, and processing the battery data to obtain a battery group total score; acquiring charging data from the vehicle non-image data, and processing the charging data to obtain a battery field; acquiring preset component information from the vehicle non-image data, and processing the preset component information to obtain a comprehensive vehicle condition score; acquiring speed information from the vehicle non-image data, and calculating the speed information to obtain a comprehensive driving behavior score; one or more of the battery group total score, the battery field, the comprehensive vehicle condition score and the comprehensive driving behavior score are used as the to-be-processed data features.

3. The method of claim 1, wherein, The inputting of the vehicle image data into a visual language large model to obtain an image correction factor feature vector comprises the following steps: encoding the vehicle image data and a preset prompt word based on an image encoder and a text encoder of the visual language large model to obtain an image vector and a text vector; splicing the image vector and the text vector to obtain a spliced vector, and processing the spliced vector based on a deep learning module of the visual language large model to obtain an output vector; processing the output vector based on a multi-task output module of the visual language large model to obtain the image correction factor feature vector.

4. The method of claim 1, wherein, Before the acquiring of the multi-source heterogeneous data of the vehicle, the method further comprises the following steps: acquiring a multi-source heterogeneous data sample of a vehicle; wherein the multi-source heterogeneous data sample comprises a vehicle non-image data sample, and the multi-source heterogeneous data sample has a residual value label; processing the vehicle non-image data sample to obtain to-be-processed data feature samples, and inputting the to-be-processed data feature samples into the at least two pre-constructed basic models respectively to obtain initial training results output by each of the pre-constructed basic models; splicing the initial training results output by each of the pre-constructed basic models into a stacked training feature vector, and inputting the stacked feature vector into a pre-constructed integrated model for processing to obtain a basic residual value training result; The first training loss value is calculated based on the basic residual value training result and the residual value label, and the first training loss value is recalculated after adjusting the model parameters of the pre-constructed integrated model when the first training loss value is greater than a preset first loss threshold, until the first training loss value is less than or equal to the preset first loss threshold, and the integrated model is obtained.

5. The method of claim 1, wherein, Before the vehicle image data is input into the visual language large model for processing to obtain an image correction factor feature vector, the method further comprises: obtaining a plurality of source heterogeneous data samples of a vehicle; wherein the plurality of source heterogeneous data samples comprise vehicle non-image data samples and vehicle image data samples, and the plurality of source heterogeneous data samples have residual value labels; inputting the vehicle non-image data samples corresponding to the to-be-processed data feature samples into the at least two basic models respectively to obtain initial training results output by each of the basic models; splicing the initial training results output by each of the basic models into a stacked training feature vector, and inputting the stacked training feature vector into the integrated model for processing to obtain a basic residual value training result; inputting the vehicle image data samples into a visual language large model to obtain an image correction factor training feature vector; splicing the basic residual value training result and the image correction factor training feature vector as an input training feature vector to input into a pre-constructed regression model to obtain a target residual value training result; calculating a second training loss value based on the target residual value training result and the residual value label, and recalculating the second training loss value after adjusting the model parameters of the pre-constructed regression model when the second training loss value is greater than a preset second loss threshold, until the second training loss value is less than or equal to the preset second loss threshold, and the regression model is obtained.

6. The method of claim 4, wherein, The processing of the vehicle non-image data samples to obtain the to-be-processed data feature samples comprises: obtaining a battery data sample from the vehicle non-image data samples; calculating a plurality of preset battery index parameter values based on the battery data sample; determining an initial score value of each preset battery index based on each preset battery index parameter value and a preset threshold; determining a weight of each preset battery index based on a preset fuzzy judgment matrix, and calculating a total score of a battery pack based on the initial score value of each preset battery index and the weight of each preset battery index to obtain the to-be-processed data feature sample.

7. The method of claim 4, wherein, The processing of the vehicle non-image data samples to obtain the to-be-processed data feature samples comprises: obtaining a charging data sample from the vehicle non-image data samples; when a voltage and remaining capacity curve graph is obtained based on a preset spline interpolation method and the charging data sample, determining a charging capacity and a power change amount based on the voltage and remaining capacity curve graph, and determining a first battery field based on the charging capacity and the power change amount as the to-be-processed data feature sample; or, When the voltage and the remaining power curve graph cannot be obtained based on the preset spline interpolation method and the charging data sample, the power difference and the power charging change are obtained by calculating based on the charging data sample, and the second battery field is determined as the to-be-processed data feature sample based on the power difference and the power charging change.

8. The method of claim 4, wherein, The processing of the vehicle non-image data sample to obtain the to-be-processed data feature sample comprises: obtaining a preset component information sample from the vehicle non-image data sample, and classifying based on the preset component information sample to obtain each preset component level; determining each preset component score value based on the each preset component level and a preset scoring strategy; calculating based on the each preset component score value and each preset component weight to obtain a comprehensive vehicle condition score as the to-be-processed data feature sample.

9. The method of claim 4, wherein, The processing of the vehicle non-image data sample to obtain the to-be-processed data feature sample comprises: obtaining a speed information sample from the vehicle non-image data sample, and calculating based on the speed information sample to determine driving data in a target time range; determining a plurality of driving parameters based on the driving data, and calculating based on each driving parameter and a preset parameter weight to obtain a comprehensive driving behavior score as the to-be-processed data feature sample.

10. The method of claim 5, wherein, The vehicle image data sample is specifically an image label, and the pre-constructed visual language large model comprises an image encoder, a text encoder, a deep learning module and a multi-task output module, and the method further comprises: inputting the vehicle image data sample and a preset prompt word into the pre-constructed visual language large model, encoding the vehicle image data sample based on the image encoder to obtain an image vector sample, and encoding the preset prompt word based on the text encoder to obtain a text vector sample; splicing the image vector and the text vector to obtain a spliced vector sample, and processing the spliced vector based on the deep learning module to obtain an output vector sample; inputting the output vector sample into the multi-task output module to obtain a plurality of task result samples, determining each task loss value based on the plurality of task result samples and the image label, and determining a total loss value based on each task loss value and a preset task weight information; when the total loss value is greater than a preset total loss threshold, adjusting the model parameters of the pre-constructed visual language large model to recalculate the total loss value until the total loss value is less than or equal to the preset total loss threshold, thereby obtaining the pre-trained visual language large model.

11. A vehicle residual value prediction device characterized by comprising: The device comprises: a first obtaining module configured to obtain multi-source heterogeneous data of a vehicle; wherein the multi-source heterogeneous data comprises vehicle non-image data and vehicle image data; a first processing module configured to process the vehicle non-image data to obtain a to-be-processed data feature; a second processing module configured to input the to-be-processed data feature into at least two basic models for processing to obtain at least two initial prediction results; The third processing module is configured to input the at least two initial prediction results into a set model for processing to obtain a basic residual value prediction result. The fourth processing module is configured to input the vehicle image data into a visual language large model for processing to obtain an image correction factor feature vector. The fifth processing module is configured to splice the basic residual value prediction result and the image correction factor feature vector as an input feature vector to input into a regression model for processing to obtain a target residual value prediction result of the vehicle.

12. An electronic device, comprising: Comprise: A processor and a memory; The processor is configured to execute the vehicle residual value prediction method according to any one of claims 1 to 10 by calling programs or instructions stored in the memory.

13. A computer-readable storage medium, characterized in that, The computer readable storage medium stores programs or instructions, which cause the computer to execute the vehicle residual value prediction method according to any one of claims 1 to 10. The computer readable storage medium stores programs or instructions, which cause the computer to execute the vehicle residual value prediction method according to any one of claims 1 to 10.

Images