Vehicle identification method based on large language model

Abstract

The invention discloses a vehicle identification method based on a language large model, and relates to the technical field of artificial intelligence and machine learning, and the main scheme is as follows: obtaining multiple items of data of a recorded vehicle; calculating a vehicle contour evaluation value and a vehicle appearance evaluation value, and further obtaining a comprehensive matching index; calculating a vehicle refitting discrimination index, and further calculating a vehicle configuration level discrimination index; calculating a basic identification evaluation value, and further calculating a vehicle matching index of each recorded vehicle; constructing a large language model, and inputting multiple items of data of a to-be-identified vehicle into the model to obtain a vehicle matching index of the to-be-identified vehicle; calculating a similarity difference value, finding out a minimum value of the similarity difference value, and finally determining a filing vehicle specifically corresponding to the to-be-identified vehicle; calculating a model adjustment comprehensive evaluation value to obtain a model comprehensive performance index, and presetting a threshold to judge a model performance grade; according to the method, the vehicle can be accurately identified, and the accuracy and comprehensiveness of vehicle identification are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence and machine learning, and specifically to a vehicle recognition method based on a large language model. Background Art

[0002] In the modern automotive field, the evaluation and matching of vehicles is a multi-dimensional and complex task. Vehicles themselves have rich characteristics, covering aspects such as appearance, operating status, configuration, etc., and are also affected by external environments and emerging technologies such as large language models. For example, in the used car trading, it is necessary to accurately identify and evaluate vehicle performance, and price and match in combination with the situation of recorded vehicles. The auto insurance industry formulates insurance rates, processes claims, and conducts anti-fraud detection by accurately identifying and evaluating vehicle performance. How to accurately and comprehensively identify and evaluate vehicle performance and effectively match it with recorded vehicles to meet diverse needs is the key problem currently faced.

[0003] In the prior art, in terms of vehicle appearance evaluation, manual visual inspection or simple image recognition algorithms are used to judge the integrity of the vehicle appearance; while in vehicle matching, rough matching is carried out based on basic information such as vehicle brand, model, and production year. These technologies provide a certain reference basis for vehicle-related work in their respective dimensions, but there are certain limitations.

[0004] However, there are many deficiencies in the prior art. First, there is a lack of comprehensive consideration of multi-dimensional factors of vehicles, and each evaluation dimension is isolated from each other, unable to form a complete evaluation system. For example, the relationship between large language models and vehicle evaluation is not fully integrated, and the performance and potential of vehicles in intelligent scenarios cannot be accurately evaluated. In addition, the prior art lacks intelligence and precision in vehicle matching, and cannot efficiently find the vehicle most similar to the vehicle to be recognized from a large amount of recorded vehicle data, resulting in inaccurate and untimely matching results, and it is difficult to meet the high requirements for vehicle evaluation and matching in practical applications. Summary of the Invention

[0005] (I) Technical Problems to be Solved: In view of the deficiencies of the prior art, the present invention provides a vehicle recognition method based on a large language model, which includes obtaining multiple data of the vehicles on record; calculating the vehicle contour evaluation value and the vehicle appearance evaluation value, and further obtaining the comprehensive matching index; calculating the vehicle modification discrimination index, and then calculating the vehicle configuration level discrimination index; calculating the basic recognition evaluation value, and further calculating the vehicle matching index of each vehicle on record; constructing a large language model, inputting the multiple data of the vehicle to be recognized into the model to obtain the vehicle matching index of the vehicle to be recognized; calculating the similarity difference, finding the minimum value of the similarity difference, and finally determining the specific vehicle on record corresponding to the vehicle to be recognized; calculating the comprehensive evaluation value of model adjustment, and then obtaining the comprehensive performance index of the model, and presetting a threshold value to judge the performance level of the model, thus solving the problems of inaccurate recognition and low real-time performance of vehicle recognition.

[0006] (2) Technical solution: To achieve the above objectives, the present invention is realized through the following technical solutions: A vehicle recognition method based on a large language model, including: Obtaining the vehicle appearance parameter data, vehicle basic parameter data, and technical parameter data of the vehicles on record; Based on the vehicle appearance parameter data, calculating the vehicle contour evaluation value SC and the vehicle appearance evaluation value SD respectively; according to the vehicle contour evaluation value SC and the vehicle appearance evaluation value SD, calculating the comprehensive matching index SAD; based on the comprehensive matching index SAD, calculating the vehicle modification discrimination index MO, and according to the vehicle modification discrimination index MO and the vehicle basic parameter data, calculating the vehicle configuration level discrimination index LE; according to the vehicle basic parameter data and the vehicle configuration level discrimination index LE, calculating the basic recognition evaluation value PK; According to the basic recognition evaluation value PK and the vehicle basic parameter data, calculating the vehicle matching index CBI of each vehicle on record; using the vehicle appearance parameter data and the vehicle basic parameter data of each vehicle on record as input features, and using the vehicle matching index CBI of the corresponding vehicle as the output feature, constructing a large language model; inputting the vehicle appearance parameter data and the vehicle basic parameter data of the vehicle to be recognized into the large language model to obtain the vehicle matching index CBO of the vehicle to be recognized; Comparing the vehicle matching index CBO of the vehicle to be recognized with the vehicle matching index CBI of the vehicle on record, calculating the similarity difference d; finding the minimum value d of the similarity difference min and then determining the specific vehicle on record corresponding to the vehicle to be recognized; Based on the technical parameter data, calculating the comprehensive evaluation value LPE of model adjustment; according to the comprehensive evaluation value LPE of model adjustment and the technical parameter data, calculating the comprehensive performance index MPI of the model; presetting the threshold value of the comprehensive performance index of the model, comparing the comprehensive performance index MPI of the model with the threshold value of the comprehensive performance index of the model, and judging the performance level of the model according to the comparison result.

[0007] In the preferred solution of the above vehicle recognition method based on a large language model: The method for calculating the vehicle contour evaluation value SC is as follows: The vehicle appearance parameter data includes the vehicle length LP, width WP, height HP, and the inclination angle EP of the vehicle's front windshield; Based on the length L, width W, height H, and the inclination angle E of the vehicle's front windshield, the formula for calculating the vehicle contour evaluation value SC is: ; where LH is the standard length of the same type of vehicle; WH is the standard width of the same type of vehicle; HH is the standard height of the same type of vehicle; EH is the standard inclination angle of the same type of vehicle; K1 is the weight coefficient of, with a value range of 0.2 - 0.3; K2 is the weight coefficient of, with a value range of 0.2 - 0.4; K3 is the weight coefficient of, with a value range of 0.3 - 0.4; K4 is the weight coefficient of, with a value range of 0.1 - 0.3; and K1 + K2 + K3 + K4 = 1.

[0008] In the preferred solution of the above vehicle recognition method based on a large language model: The method for calculating the vehicle appearance evaluation value SD is as follows: The vehicle appearance parameter data further includes the logo area A of the vehicle, the curvature K of the body lines, the number N of the wheel hub spokes, and the logo color coefficient C; Based on the logo area A, the curvature K of the body lines, and the number N of the wheel hub spokes, the formula for calculating the vehicle appearance evaluation value SD is: ; where AS is the standard logo area; KS is the standard curvature of the body lines; NS is the standard number of the wheel hub spokes; ω1 is the weight coefficient of, with a value range of 0.1 - 0.3; ω2 is the weight coefficient of, with a value range of 0.3 - 0.4; ω3 is the weight coefficient of, with a value range of 0.3 - 0.5; and ω1 + ω2 + ω3 = 1.

[0009] In the preferred solution of the above vehicle recognition method based on a large language model: The method for calculating the comprehensive matching index SAD is as follows: Based on the vehicle contour evaluation value SC and the vehicle appearance evaluation value SD, the formula for calculating the comprehensive matching index SAD is: ; Among them, φ1 is the weight coefficient of the vehicle contour evaluation value SC, and its value ranges from 0.5 to 0.7; φ2 is the weight coefficient of the vehicle appearance evaluation value SD, and its value ranges from 0.3 to 0.5; and φ1 + φ2 = 1.

[0010] In the preferred scheme of the above vehicle recognition method based on a language large model: The method for calculating the vehicle configuration level discrimination index LE is as follows: The vehicle basic parameter data includes the engine power PI of the vehicle; Based on the comprehensive matching index SAD, the formula for calculating the vehicle modification discrimination index MO is as follows: ; Among them, YH is the vehicle usage years; According to the vehicle modification discrimination index MO and the engine power PI, the formula for calculating the vehicle configuration level discrimination index LE is as follows: ; Among them, WP is the maximum load capacity of the vehicle.

[0011] In the preferred scheme of the above vehicle recognition method based on a language large model: The method for calculating the basic recognition evaluation value PK is as follows: The vehicle basic parameter data also includes the total driving mileage MG of the recorded vehicle; According to the total driving mileage MG and the vehicle configuration level discrimination index LE, the formula for calculating the basic recognition evaluation value PK is as follows: ; Among them, MC is the average maintenance cost per 100 kilometers.

[0012] In the preferred scheme of the above vehicle recognition method based on a language large model: The method for calculating the vehicle matching index CBI of each recorded vehicle is as follows: The vehicle basic parameter data also includes the average annual driving mileage AMM and the fuel consumption per 100 kilometers FC of the recorded vehicle; According to the basic recognition evaluation value PK, the average annual driving mileage AMM and the fuel consumption per 100 kilometers FC, the formula for calculating the vehicle matching index CBI of each recorded vehicle is as follows: ; Among them, FP is the fuel price; IC is the vehicle insurance cost; DR is the vehicle depreciation rate.

[0013] In the preferred scheme of the above vehicle recognition method based on a language large model: The method for determining the recorded vehicle corresponding to the vehicle to be recognized is as follows: According to the vehicle matching index CBO of the vehicle to be recognized and the vehicle matching index CBI of each registered vehicle, calculate the similarity difference d between the vehicle matching index CBO of the vehicle to be recognized and the vehicle matching index CBI of each registered vehicle. The formula is as follows: ; By comparing all the similarity difference d values, find the minimum similarity difference value d min , and the registered vehicle corresponding to it is the most similar to the vehicle to be recognized, and then determine the specific registered vehicle corresponding to the vehicle to be recognized.

[0014] In the above preferred scheme of a vehicle recognition method based on a language model: The method for calculating the comprehensive evaluation value LPE of model adjustment is as follows: The technical parameter data also includes the number of training rounds AC of the language model, the hidden layer dimension HC of the language model, the number of attention heads EC of the language model, and the word embedding dimension DC of the language model; According to the number of training rounds AC of the language model, the hidden layer dimension HC of the language model, the number of attention heads EC of the language model, and the word embedding dimension DC of the language model, calculate the comprehensive evaluation value LPE of model adjustment. The calculation formula is as follows: ; Among them, COR is the vehicle recognition accuracy rate.

[0015] In the above preferred scheme of a vehicle recognition method based on a language model: The method for judging the model performance level is as follows: The technical parameter data also includes the training duration IM of the language model and the update period FHJ of the language model; According to the comprehensive evaluation value LPE of model adjustment, the training duration IM of the language model, and the update period FHJ of the language model, calculate the comprehensive performance index MPI of the model. The calculation formula is as follows: ; Among them, PS is the parameter order of magnitude of the language model; DW is the proportion of vehicle-related information covered in the training data of the language model; The comprehensive performance index threshold of the model includes the comprehensive performance index threshold one MPI1 and the comprehensive performance index threshold two MPI2, and MPI1 < MPI2; When MPI ≤ MPI1, it is judged that the model performance is poor; When MPI1 < MPI ≤ MPI2, it is judged that the model performance is medium; When MPI > MPI2, it is judged that the model performance is good.

[0016] (III) Beneficial effects: The present invention provides a vehicle recognition method based on a large language model, which has the following beneficial effects: (1) The vehicle appearance parameter data can visually present the external characteristics of the vehicle, facilitating quick recognition and differentiation, and improving efficiency and accuracy. The vehicle basic parameter data provides a solid basis for comprehensively evaluating the vehicle performance, determining the applicable scenarios, etc., and helps with accurate classification and matching. The technical parameter data can deeply explore the potential and adaptability of the vehicle in terms of intelligence.

[0017] (2) Calculate the vehicle contour evaluation value and the vehicle appearance evaluation value, and then further obtain the comprehensive matching index, as well as the subsequent vehicle modification discrimination index and vehicle configuration level discrimination index. This series of shape-based evaluations and calculations can meticulously analyze the vehicle's appearance characteristics and modification situation, accurately locate the vehicle's configuration level, provide a key shape basis for the accurate recognition and classification of vehicles, and improve the pertinence and accuracy of vehicle recognition.

[0018] (3) Calculating the matching index of the recorded vehicles can comprehensively consider various characteristics of the vehicles, making the evaluation more scientific and objective. Constructing a large language model with various parameter data as input features gives full play to the value of the data, enabling the model to deeply explore the potential correlations between the data and improving the accuracy of vehicle matching. Inputting the data of the vehicle to be recognized into the model to obtain the matching index greatly improves the efficiency of vehicle recognition and matching.

[0019] (4) Calculating the similarity difference can quickly and accurately locate the vehicle that best matches it among numerous recorded vehicles, greatly improving the accuracy and efficiency of vehicle recognition.

[0020] (5) Calculating the model adjustment comprehensive evaluation value and the model comprehensive performance index can comprehensively and objectively measure the model performance, avoiding subjective assumptions. Presetting the model comprehensive performance index threshold provides a clear standard for judging the model performance level, making the evaluation results consistent and comparable. Knowing the model performance level in a timely manner based on the comparison results helps quickly locate the models with poor performance, and make targeted optimization adjustments or updates, thereby improving the overall quality and efficiency of the models. Description of the Drawings

[0021] Figure 1 It is a step schematic diagram of a vehicle recognition method based on a large language model of the present invention. Detailed Embodiment

[0022] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0023] Please refer to Figure 1 , the present invention provides a vehicle recognition method based on a large language model, including: Step 1: Obtain the vehicle appearance parameter data, vehicle basic parameter data, and technical parameter data of the vehicles on record.

[0024] Comprehensive Step 1: The vehicle appearance parameter data can intuitively present the external state of the vehicle, facilitating the identification and differentiation of different vehicles. The vehicle basic parameter data covers the basic information of the vehicle, helping to comprehensively understand the basic performance and specifications of the vehicle, and providing a basic basis for the evaluation, classification, and matching of the vehicle. The technical parameter data delves into the technical level of the vehicle and can accurately evaluate the technical level and functional characteristics of the vehicle.

[0025] Step 2: Calculate the vehicle contour evaluation value SC and the vehicle appearance evaluation value SD based on the vehicle appearance parameter data respectively; calculate the comprehensive matching index SAD according to the vehicle contour evaluation value SC and the vehicle appearance evaluation value SD; calculate the vehicle modification discrimination index MO based on the comprehensive matching index SAD, and calculate the vehicle configuration level discrimination index LE according to the vehicle modification discrimination index MO and the vehicle basic parameter data; calculate the basic recognition evaluation value PK according to the vehicle basic parameter data and the vehicle configuration level discrimination index LE.

[0026] Step 201: Calculate the vehicle contour evaluation value SC, and the specific method is as follows: The vehicle appearance parameter data includes the length LP, width WP, height HP of the vehicle, and the inclination angle EP of the vehicle front windshield.

[0027] It should be noted that the vehicle length LP, width WP, and height HP are obtained by actually measuring the vehicle using tools such as a tape measure and a laser rangefinder. The inclination angle of the vehicle front windshield is directly measured using an angle measuring instrument to obtain the inclination angle EP of the vehicle front windshield.

[0028] According to the length L, width W, height H, and the inclination angle E of the vehicle front windshield, calculate the vehicle contour evaluation value SC, and the formula used is: ; Among them, LH is the standard length of vehicles of the same type; WH is the standard width of vehicles of the same type; HH is the standard height of vehicles of the same type; EH is the standard inclination angle of vehicles of the same type; K1 is the weight coefficient of , with a value range of 0.2 to 0.3; K2 is the weight coefficient of , with a value range of 0.2 to 0.4; K3 is

[0029] the weight coefficient of

[0030] It should be noted that by visiting the official website of the automobile manufacturer and looking up the detailed information of the corresponding model on the technical specifications or product information page, the standard length LH of vehicles of the same type, the standard width WH of vehicles of the same type, the standard height HH of vehicles of the same type, and the standard inclination angle EH of vehicles of the same type can be obtained. 、 、 and calculate the absolute value of the difference between the vehicle and the standard value of vehicles of the same type; the absolute values of these differences reflect the degree of deviation of the actual size and angle of the vehicle from the standard value. Multiplying each by its respective weight coefficient indicates the degree of emphasis on different parameters when calculating the vehicle profile evaluation value SC. By multiplying each parameter by its corresponding weight coefficient and then summing, the total deviation degree index is obtained. Finally, subtracting this total deviation degree index from 1 gives the vehicle profile evaluation value SC. The closer the value of SC is to 1, the closer the vehicle profile is to the standard profile of vehicles of the same type, and the smaller the deviation degree; the smaller the value of SC, the greater the difference between the vehicle profile and the standard profile.

[0031] Step 202: Calculate the vehicle appearance evaluation value SD, and the specific method is as follows: The vehicle appearance parameter data also includes the logo area A of the vehicle, the body line curvature K, the number of wheel hub spokes N, and the logo color coefficient C.

[0032] It should be noted that the logo area A of the vehicle can be obtained by taking a photo of the vehicle logo, then setting the scale in image measurement software such as ImageJ and Adobe Photoshop, using the area measurement tool in the image measurement software to outline the contour of the logo, and the software can automatically calculate the area of the logo to obtain the logo area A of the vehicle.

[0033] By using a 3D laser scanner to scan the vehicle body, 3D point cloud data is obtained. Then, 3D modeling software such as CATIA and Geomagic is used to process and model the 3D point cloud data. In the modeling software, by selecting multiple points on the body line, the curve equation of the body line is fitted, and then the curvature is calculated according to the curve equation. For example, in CATIA, the curve fitting tool can be used to fit the waist line on the side of the body, and then the curvature value of this line can be obtained through the curvature analysis function of the software, so as to obtain the body line curvature K.

[0034] By observing the vehicle's wheels and viewing them from multiple angles such as the front and side, the number of spokes on the wheels is output to obtain the number of wheel spokes N.

[0035] By using a spectrophotometer or color analyzer to align with the logo, the instrument will measure data such as the spectral reflectance of the logo color. The instrument will output the measurement results in the form of parameters in multiple color spaces, such as RGB, Lab, XYZ, etc. And these parameter values are converted into grayscale values, and the grayscale values are used as the logo color coefficient C. For example, in the RGB color space, when R = 120, G = 80, B = 50, the grayscale value = 0.299×120 + 0.587×80 + 0.114×50 = 83.7, then the logo color coefficient C is 83.7.

[0036] According to the logo area A, the body line curvature K, and the number of wheel spokes N, the vehicle appearance evaluation value SD is calculated, and the formula is as follows: ; where, AS is the standard logo area; KS is the standard body line curvature; NS is the standard number of wheel spokes; ω1 is the weight coefficient of, with a value range of 0.1 to 0.3; ω2 is the weight coefficient of, with a value range of 0.3 to 0.4; ω3 is the weight coefficient of, with a value range of 0.3 to 0.5; and ω1 + ω2 + ω3 = 1.

[0037] It should be noted that to obtain the standard logo area AS, since automobile manufacturers usually have a set of standard design specification documents for each vehicle model during the vehicle design and production process. Therefore, the standard logo area AS of this vehicle model can be obtained through the official technical materials of the automobile manufacturer, vehicle design manuals, etc.

[0038] Obtain the standard body line curvature KS. Since the design team of the automobile manufacturer will use CAD software to draw the three-dimensional model and detailed design drawings of the vehicle when designing the vehicle. These drawings will contain the precise geometric information of the body lines. The design drawings can be obtained from the design department of the automobile manufacturer or through legal channels, and the standard body line curvature KS can be extracted from them.

[0039] Obtain the standard number of wheel spokes NS. The official documents such as the vehicle purchase invoice, vehicle certificate of conformity, and user manual of the vehicle usually contain the detailed configuration information of the vehicle, which will clearly mark the wheel styles equipped with the vehicle model and the corresponding number of spokes, and the standard number of wheel spokes NS can be obtained from them.

[0040] It should be noted that the numerator is 1, indicating that the change of the denominator is used to reflect the impact of the difference between the vehicle appearance parameters and the standard value on the appearance valuation. 、 and These three items respectively represent the squares of the deviations of the vehicle logo area, body line curvature, and number of wheel spokes from their respective standard values; and then multiply by their respective weight coefficients respectively, which ensures the integrity of the proportions of the three weight coefficients in the overall calculation. They jointly determine the comprehensive impact of the three parameter deviations on the denominator. To sum up, the formula comprehensively considers the impacts of the three factors of the logo area, body line curvature, and number of wheel spokes on the vehicle appearance, and calculates the appearance valuation of the vehicle according to their deviations from the standard values and their respective weights, providing a quantitative method for evaluating the vehicle appearance.

[0041] Step 203: Calculate the comprehensive matching index SAD. The specific method is as follows: Calculate the comprehensive matching index SAD according to the vehicle contour evaluation value SC and the vehicle appearance evaluation value SD. The calculation formula is: ; where, φ1 is the weight coefficient of the vehicle contour evaluation value SC, and its value ranges from 0.5 to 0.7; φ2 is the weight coefficient of the vehicle appearance evaluation value SD, and its value ranges from 0.3 to 0.5; and φ1 + φ2 = 1.

[0042] It should be noted that this formula comprehensively considers the two important parameters of the vehicle contour evaluation value SC and the vehicle appearance evaluation value SD. Each parameter has a corresponding weight coefficient, indicating the degree of emphasis on different parameters when calculating the comprehensive matching index SAD. By multiplying each parameter by its corresponding weight coefficient and then summing, a comprehensive matching index SAD that can comprehensively describe the vehicle is obtained.

[0043] Step 204: Calculate the vehicle configuration level discrimination index LE. The specific method is as follows: The vehicle basic parameter data includes the engine power PI of the vehicle.

[0044] It should be noted that the engine power PI can be obtained through information such as the real-time power or maximum power of the engine displayed on the vehicle's instrument panel or the vehicle computer display screen.

[0045] Based on the comprehensive matching index SAD, the vehicle modification discrimination index MO is calculated, and the formula is as follows: ; Among them, YH is the vehicle usage years.

[0046] It should be noted that to obtain the vehicle usage years YH, by checking the vehicle nameplate on the vehicle, the manufacturing date of the vehicle can be obtained, and based on the manufacturing date and the current date, the vehicle usage years can be obtained.

[0047] It should be noted that indicates that it has a greater contribution to the overall result when the vehicle usage years are longer. indicates that it comprehensively considers the impacts of SAD and YH on the vehicle modification discrimination index MO from different degrees and weights. This term indicates that it can be dynamically adjusted within a large range according to the change of SAD, thus affecting the value of the entire fraction. indicates that it plays a role in fine-tuning the denominator value in the denominator, and together with the previous part of the denominator, further refines the change trend of the denominator. This formula comprehensively considers the impacts of the comprehensive matching index SAD and the vehicle usage years YH on the vehicle modification discrimination index MO.

[0048] According to the vehicle modification discrimination index MO and the engine power PI, the vehicle configuration level discrimination index LE is calculated, and the calculation formula is: ; Among them, WP is the maximum load capacity of the vehicle.

[0049] It should be noted that to obtain the maximum load capacity WP of the vehicle, by checking the vehicle nameplate on the vehicle, the maximum load capacity of the vehicle can be obtained.

[0050] It should be noted that This part considers the impacts of the vehicle engine power PI and the maximum load capacity WP of the vehicle. As the engine power and the maximum load capacity increase, the value of this term will gradually increase, so that this part of the numerator has a greater contribution to the overall result when the engine power and the maximum load capacity are large. This part and the previous part together constitute the numerator, and comprehensively consider the impacts of MO, PI and WP on the vehicle configuration level discrimination index LE from different degrees and weights. This part is based on the cube operation of MO. The coefficient 3 is used to adjust the weight of this part in the numerator, which can reasonably reflect the influence of vehicle modification on the configuration level discrimination index LE according to the actual situation. This term increases as MO increases. It further fine-tunes the denominator value in the denominator, making the change of the denominator more complex and diverse.

[0051] Comprehensive steps 201 to 204: Through the analysis and calculation of vehicle shape parameter data, accurately grasp the vehicle contour and appearance characteristics, and then accurately identify the vehicle modification situation and configuration level, providing a key basis for vehicle identification, avoiding misjudgment caused by similar appearance or modification, and improving the accuracy and professionalism of vehicle identification.

[0052] Step 3: According to the basic recognition evaluation value PK and vehicle basic parameter data, calculate the vehicle matching index CBI for each registered vehicle; use the vehicle appearance parameter data and vehicle basic parameter data of each registered vehicle as input features, and use the vehicle matching index CBI of the corresponding vehicle as the output feature to construct a language large model; input the vehicle appearance parameter data and vehicle basic parameter data of the vehicle to be identified into the language large model to obtain the vehicle matching index CBO of the vehicle to be identified.

[0053] Step 301: Calculate the basic recognition evaluation value PK. The specific method is as follows: The vehicle basic parameter data also includes the total driving mileage MG of the registered vehicle.

[0054] It should be noted that by using an OBD diagnostic instrument to connect to the OBD interface of the vehicle and reading the driving mileage information stored in the vehicle's electronic control unit through the operation interface of the diagnostic instrument, the total driving mileage MG of the registered vehicle can be obtained.

[0055] According to the total driving mileage MG and the vehicle configuration level discrimination index LE, calculate the basic recognition evaluation value PK. The formula is: ; Among them, MC is the average maintenance cost per 100 kilometers.

[0056] It should be noted that to obtain the average maintenance cost per 100 kilometers MC, obtain the historical maintenance records of the vehicle through the vehicle maintenance manual, and further obtain the total maintenance cost SS. According to the total driving mileage MG and the total maintenance cost SS, calculate the average maintenance cost per 100 kilometers MC, MC = SS / MG × 100.

[0057] It should be noted that The total driving mileage is scaled. Dividing by 10,000 is to adjust the value of the driving mileage to an appropriate magnitude so that it can be calculated on the same scale as other parameters, avoiding an excessive impact on the overall result due to the overly large value of the driving mileage. This part indicates that when the vehicle age increases, the denominator increases, and the value of the entire fraction decreases, thus restricting the basic recognition evaluation value. This part indicates that when the average maintenance cost per 100 kilometers increases, the denominator increases, and the value of the fraction decreases, thus having a negative impact on the basic recognition evaluation value. The formula combines factors such as the vehicle configuration level discrimination index, total driving mileage, vehicle age, and average maintenance cost per 100 kilometers to calculate the basic recognition evaluation value PK.

[0058] Step 302: Calculate the vehicle matching index CBI for each registered vehicle. The specific method is as follows: The vehicle basic parameter data also includes the average annual driving mileage AMM and fuel consumption per 100 kilometers FC of the registered vehicle.

[0059] It should be noted that by checking the monthly driving distance data recorded in the vehicle information system and then adding these data and dividing by 12, the average annual driving mileage AMM of the vehicle can be obtained.

[0060] During the vehicle R & D process, using a professional chassis dynamometer and fuel consumption test equipment to test the vehicle according to the WLTP test cycle, the fuel consumption per 100 kilometers FC of the vehicle can be obtained.

[0061] According to the basic recognition evaluation value PK, average annual driving mileage AMM, and fuel consumption per 100 kilometers FC, calculate the vehicle matching index CBI for each registered vehicle. The formula is as follows: ; Among them, FP is the fuel price; IC is the vehicle insurance cost; DR is the vehicle depreciation rate.

[0062] It should be noted that to obtain the fuel price FP, by visiting the fuel price monitoring website and entering the name of the city where you are located, the website will display the fuel price in that city, and thus obtain the fuel price FP.

[0063] To obtain the vehicle insurance cost IC, by checking the vehicle insurance policy, directly obtain the vehicle insurance cost IC.

[0064] To obtain the vehicle depreciation rate DR, by referring to the vehicle depreciation rate reference table in the automotive industry, find the vehicle age and vehicle brand of this vehicle, and refer to the depreciation rate table to obtain the vehicle depreciation rate DR.

[0065] It should be noted that Annual average mileage, fuel consumption per 100 kilometers, and fuel price are considered, which reflect the factors related to vehicle usage costs. The longer the mileage, the lower the fuel consumption, and the lower the fuel price, the greater the contribution of this part to the matching index. The non-linear effects of annual average mileage and basic recognition evaluation value are comprehensively considered. When the mileage and basic recognition evaluation value are large, the value of this part will gradually increase, but the growth rate will gradually slow down to avoid over-amplifying its impact. The importance of the basic recognition evaluation value is emphasized, while avoiding its excessive influence on the matching index. It means that the higher the vehicle insurance cost, the larger the denominator, which restricts the vehicle matching index. It means that the higher the vehicle depreciation rate, the larger the denominator, and the greater the negative impact on the vehicle matching index. This formula comprehensively considers factors such as the basic recognition evaluation value, usage cost, insurance cost, and depreciation rate of the vehicle to calculate the vehicle matching index, and each part interacts with each other to more comprehensively evaluate the matching degree of the vehicle.

[0066] Taking the vehicle appearance parameter data, vehicle basic parameter data, and technical parameter data of each recorded vehicle as input features, and taking the vehicle matching index CBI of the corresponding vehicle as the output feature, a language large model is constructed; the vehicle appearance parameter data, vehicle basic parameter data, and technical parameter data of the vehicle to be identified are input into the language large model to obtain the vehicle matching index CBO of the vehicle to be identified.

[0067] Comprehensive steps 301 - 302: This method organically combines various types of parameter data of the vehicle, processes them through a language large model, fully explores the potential value of the data, and provides a more comprehensive and accurate basis for vehicle matching. With the powerful learning and generalization ability of the language large model, the vehicle matching index is calculated more accurately, improving the accuracy and reliability of the matching.

[0068] Step Four: Compare the vehicle matching index CBO of the vehicle to be identified with the vehicle matching index CBI of the recorded vehicles, calculate the similarity difference d; find the minimum similarity difference d min , and then determine the specific recorded vehicle corresponding to the vehicle to be identified.

[0069] Step 401: Determine the specific recorded vehicle corresponding to the vehicle to be identified. The specific method is as follows: According to the vehicle matching index CBO of the vehicle to be identified and the vehicle matching index CBI of each recorded vehicle, calculate the similarity difference d between the vehicle matching index CBO of the vehicle to be identified and the vehicle matching index CBI of each recorded vehicle. The formula is: ; By comparing all the similarity difference d values, find the minimum similarity difference value d among them min , and the archived vehicle corresponding to it is the most similar to the vehicle to be identified, thereby determining the specific archived vehicle corresponding to the vehicle to be identified.

[0070] Comprehensive step four: This method avoids the subjectivity and low efficiency of traditional manual identification, can quickly and accurately find the vehicle most similar to it among many archived vehicles, and greatly improves the accuracy and speed of vehicle identification.

[0071] Step five: Calculate the comprehensive evaluation value LPE for model adjustment based on the technical parameter data; calculate the comprehensive performance index MPI of the model according to the comprehensive evaluation value LPE for model adjustment and the technical parameter data; preset the threshold of the comprehensive performance index of the model, compare the comprehensive performance index MPI of the model with the threshold of the comprehensive performance index of the model, and judge the performance level of the model according to the comparison result.

[0072] Step 501: Calculate the comprehensive evaluation value LPE for model adjustment. The specific method is as follows: The technical parameter data also includes the number of training rounds AC of the language large model, the hidden layer dimension HC of the language large model, the number of attention heads EC of the language large model, and the word embedding dimension DC of the language large model.

[0073] It should be noted that during the training process of the language large model, detailed training logs are usually recorded. These logs will contain relevant information for each round of training, including the start time, end time, training batches, loss values, etc. By analyzing these logs, the total number of training rounds can be counted. For example, the logs may record starting from the 1st round of training to the 100th round of training, then the number of training rounds AC is 100, and the number of training rounds AC of the language large model is obtained in this way.

[0074] Since the architecture of the language large model is usually defined by code or configuration files. In these files, the dimension information of each layer in the model will be clearly specified. The source code of the model or the relevant configuration files can be viewed to find the setting of the hidden layer dimension. For example, in a language large model project based on TensorFlow, the architecture of the model is defined in a file named "model_architecture.py", where there is a code line like "hidden_layer_dimension = 1024", and the value of 1024 here is the value of the hidden layer dimension HC, and the hidden layer dimension HC of the language large model is obtained in this way.

[0075] By using Netron to visualize the deep learning model, the trained model file is loaded, and then the structure of the model is displayed graphically. From this, the number of attention heads EC of the large language model can be obtained. For example, by using Netron to open the weight file of a large language model, in the visualization interface, the attention module part can be clearly seen, showing 6 attention heads. Then the EC of this model is 6.

[0076] In the code of the large language model, the word embedding layer is the part that converts the input words into vector representations. In the initialization code of the word embedding layer, the word embedding dimension will be specified. For example, in the code of a natural language processing project, it is seen that "embedding_layer=nn.Embedding(10000, 300)", and here 300 is the value of the word embedding dimension DC, indicating that the model maps each word in the vocabulary to a 300-dimensional vector space. The word embedding dimension DC of the large language model is determined according to this method.

[0077] According to the number of training rounds AC of the large language model, the hidden layer dimension HC of the large language model, the number of attention heads EC of the large language model, and the word embedding dimension DC of the large language model, calculate the model adjustment comprehensive evaluation value LPE. The calculation formula is: ; Among them, COR is the vehicle recognition accuracy rate.

[0078] It should be noted that to obtain the vehicle recognition accuracy rate COR, a vehicle on-vehicle recognition test experiment is carried out on the vehicle. According to the total number of vehicle recognition tests NTO and the results of each vehicle recognition test in the background database of the on-vehicle recognition system, the number of correctly recognized vehicle times NCO is counted, and the vehicle recognition accuracy rate COR is calculated. COR = NCO / NTO × 100%.

[0079] It should be noted that the numerator means that the structure and capabilities of the model are combined with the actual performance of the model, jointly affecting the model adjustment comprehensive evaluation value. The denominator represents a non-linear adjustment of the overall evaluation value according to the accuracy rate, so that the evaluation result can more comprehensively and accurately reflect the comprehensive performance of the model in the vehicle recognition task. This formula calculates the model adjustment comprehensive evaluation value LPE by combining the vehicle recognition accuracy rate with multiple technical parameters of the large language model.

[0080] Step 502: Judge the model performance level. The specific method is: The technical parameter data also includes the training duration IM obtained by the large language model and the update period FHJ of the large language model.

[0081] It should be noted that by viewing the log file generated during the training of the language large model in TensorFlow, the timestamps at the start and end of the training are found, and the training duration IM of the language large model is obtained by subtracting the start time from the end time.

[0082] The update cycle of the model is determined by viewing the version history. The submission time of each version and the interval between versions can reflect the update frequency of the model. For example, if it takes 30 days to update from one version to the next, then the update cycle is approximately 30 days, and in this way, the update cycle FHJ of the language large model is obtained.

[0083] According to the comprehensive evaluation value LPE of the model adjustment, the training duration IM of the language large model, and the update cycle FHJ of the language large model, the model comprehensive performance index MPI is calculated. The calculation formula is: ; Among them, PS is the parameter order of magnitude of the language large model; DW is the proportion of vehicle-related information covered in the training data of the language large model.

[0084] It should be noted that to obtain the parameter order of magnitude PS of the language large model, view the source code or architecture document of the language large model. In the definition part of the model, the number of parameters of each layer is usually specified explicitly. For example, in a language large model based on the Transformer architecture, by viewing the code of the model, it is found that there are 12 layers in the encoder part, the number of attention heads in each layer is 8, the dimension of each attention head is 64, the hidden layer dimension of the feed-forward neural network is 2048, the word embedding dimensions of the input and output are 512, etc. By calculating the number of these parameters and adding them up, the total number of parameters is about 50 million, so the parameter order of magnitude is about , and the parameter order of magnitude PS of the language large model is obtained according to this method.

[0085] To obtain the proportion DW of vehicle-related information covered in the training data of the language large model, annotate and count the training data of the language large model. First, determine the categories and features of vehicle-related information, such as vehicle brands, models, performance parameters, component names, driving-related terms, etc. Then, use an automated annotation tool to annotate each piece of data in the training data to determine whether it contains vehicle-related information. Finally, count the number of data entries containing vehicle-related information and divide it by the total number of training data entries to obtain the proportion DW of vehicle-related information covered. For example, if there are 1 million pieces of training data for the language large model, and after annotation and statistics, it is found that 100,000 pieces of data contain vehicle-related information, then the proportion DW of vehicle-related information covered = 10000 / 1000000 = 0.1.

[0086] It should be noted that This part as a whole reflects the trade - off between the training duration and the update period. That is, the longer the training duration and the relatively reasonable the update period, the larger the value of this part and the greater the positive contribution to MPI. It indicates that when the parameter order of magnitude is small, increasing the parameter order of magnitude has an obvious promoting effect on the model comprehensive performance index MPI. It shows that the higher the proportion of vehicle - related information in the training data, the greater the positive contribution to the model comprehensive performance index MPI. The entire formula combines factors such as the model adjustment comprehensive evaluation value LEP, training duration IM, parameter order of magnitude PS, update period FHJ, and the proportion DW of vehicle - related information covered in the training data in a specific functional form, comprehensively considering multiple aspects of the model, so as to calculate the model comprehensive performance index MPI.

[0087] Preset the model comprehensive performance index threshold. The model comprehensive performance index threshold includes the model comprehensive performance index threshold one MPI1 and the model comprehensive performance index threshold two MPI2, and MPI1 < MPI2; by collecting relevant data of multiple language large models, including the model adjustment comprehensive evaluation value LEP, training duration IM, update period FHJ, parameter order of magnitude PS, and the proportion DW of vehicle - related information covered in the training data, etc.; analyzing these data, calculating the model comprehensive performance index of each language large model, further calculating the average value MPIavg and standard deviation σ of the model comprehensive performance index, taking MPIavg - σ as the model comprehensive performance index threshold one MPI1; taking MPIavg + σ as the model comprehensive performance index threshold two MPI2. When MPI ≤ MPI1, it is judged that the model performance is poor and the model needs to be comprehensively updated immediately. When MPI1 < MPI ≤ MPI2, it is judged that the model performance is medium and the model needs to be adjusted. When MPI > MPI2, it is judged that the model performance is good.

[0088] Integrate step 501 to step 502: Calculating the model adjustment comprehensive evaluation value and the model comprehensive performance index can comprehensively and objectively measure the model performance. Presetting the model comprehensive performance index threshold provides a clear standard for the judgment of the model performance level, making the evaluation results consistent and comparable. Knowing the model performance level in a timely manner based on the comparison results helps to quickly locate the models with poor performance and conduct targeted optimization, adjustment or update, so as to improve the overall quality and efficiency of the model.

[0089] The above embodiments can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. Those of ordinary skill in the art will appreciate that the units and algorithm steps of the examples described in connection with the embodiments disclosed herein can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are executed in hardware or software depends on the specific application and design constraints of the technical solution.

[0090] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0091] As described above, the above are only specific embodiments of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art within the technical scope disclosed in the present application can easily think of changes or substitutions, which should all be covered by the protection scope of the present application.

Claims

1. A vehicle identification method based on a large language model, characterized in that: include: Obtain vehicle appearance parameter data, vehicle basic parameter data and technical parameter data of registered vehicles; Based on the vehicle appearance parameter data, the vehicle contour evaluation value SC and the vehicle appearance evaluation value SD are respectively calculated; based on the vehicle contour evaluation value SC and the vehicle appearance evaluation value SD, the comprehensive matching index SAD is calculated; based on the comprehensive matching index SAD, the vehicle modification discrimination index MO is calculated, and based on the vehicle modification discrimination index MO and the vehicle basic parameter data, the vehicle configuration level discrimination index LE is calculated; based on the vehicle basic parameter data and the vehicle configuration level discrimination index LE, the basic recognition evaluation value PK is calculated; According to the basic recognition evaluation value PK and the basic vehicle parameter data, the vehicle matching index CBI of each registered vehicle is calculated; the vehicle appearance parameter data and vehicle basic parameter data of each registered vehicle are used as input features, and the vehicle matching index CBI of the corresponding vehicle is used as output features to build a language model; Input the vehicle appearance parameter data and vehicle basic parameter data of the vehicle to be identified into the language model to obtain the vehicle matching index CBO of the vehicle to be identified; Compare the vehicle matching index CBO of the vehicle to be identified with the vehicle matching index CBI of the registered vehicle, calculate the similarity difference d; find the minimum similarity difference d min , and then determine the specific registered vehicle corresponding to the vehicle to be identified; Based on the technical parameter data, the model is calculated to adjust the comprehensive evaluation value LPE; Based on the model adjustment comprehensive evaluation value LPE and technical parameter data, the model comprehensive performance index MPI is calculated; A model comprehensive performance index threshold is preset, the model comprehensive performance index MPI is compared with the model comprehensive performance index threshold, and the model performance level is determined according to the comparison result.

2. The vehicle identification method based on a language large model according to claim 1, characterized in that: The method for calculating the vehicle profile evaluation value SC is: The vehicle appearance parameter data includes the vehicle's length LP, width WP, height HP and the vehicle's front windshield inclination angle EP; The vehicle profile evaluation value SC is calculated based on the length L, width W, height H and the vehicle front windshield inclination angle E, according to the formula: ； Among them, LH is the standard length of the same type of vehicle; WH is the standard width of the same type of vehicle; HH is the standard height of the same type of vehicle; EH is the standard tilt angle of the same type of vehicle; K1 is The weight coefficient is 0.2~0.3; K2 is The weight coefficient is 0.2~0.4; K3 is The weight coefficient is 0.3~0.4; K4 is The weight coefficient is between 0.1 and 0.3, and K1+K2+K3+K4=1.

3. The vehicle identification method based on a language large model according to claim 2 is characterized in that: The method for calculating the vehicle appearance evaluation value SD is: The vehicle appearance parameter data also includes the vehicle logo area A, body line curvature K, wheel hub spoke number N and logo color coefficient C; The vehicle appearance evaluation value SD is calculated based on the logo area A, the body line curvature K and the number of wheel spokes N, and the formula is: ； Among them, AS is the standard logo area; KS is the standard body line curvature; NS is the number of standard wheel spokes; ω1 is The weight coefficient is 0.1~0.3; ω2 is The weight coefficient is 0.3~0.4; ω3 is The weight coefficient is between 0.3 and 0.5, and ω1+ω2+ω3=1.

4. The vehicle identification method based on a language large model according to claim 3 is characterized in that: The method for calculating the comprehensive matching index SAD is: According to the vehicle contour evaluation value SC and the vehicle appearance evaluation value SD, the comprehensive matching index SAD is calculated, and the calculation formula is: ； Among them, φ1 is the weight coefficient of the vehicle contour evaluation value SC, which is 0.5~0.7; φ2 is the weight coefficient of the vehicle appearance evaluation value SD, which is 0.3~0.5; and φ1+φ2=1.

5. The vehicle identification method based on a language large model according to claim 4 is characterized in that: The method for calculating the vehicle configuration level judgment index LE is: The basic vehicle parameter data includes the vehicle's engine power PI; Based on the comprehensive matching index SAD, the vehicle modification identification index MO is calculated based on the formula: ； Among them, YH is the age of the vehicle; According to the vehicle modification identification index MO and the engine power PI, the vehicle configuration level identification index LE is calculated. The calculation formula is: ； Among them, WP is the maximum load capacity of the vehicle.

6. The vehicle identification method based on a language large model according to claim 5 is characterized in that: The method for calculating the basic identification evaluation value PK is: The basic vehicle parameter data also includes the total mileage MG of the registered vehicle; Based on the total driving mileage MG and the vehicle configuration level discrimination index LE, calculate the basic recognition evaluation value PK. The formula used is as follows: ； Among them, MC is the average maintenance cost per 100 kilometers.

7. The vehicle identification method based on a language large model according to claim 6 is characterized in that: The method for calculating the vehicle matching index CBI of each recorded vehicle is as follows: The vehicle basic parameter data also includes the average annual driving mileage AMM and the fuel consumption per 100 kilometers FC of the recorded vehicle; Based on the basic recognition evaluation value PK, the average annual driving mileage AMM, and the fuel consumption per 100 kilometers FC, calculate the vehicle matching index CBI of each recorded vehicle. The formula used is as follows: ； Among them, FP is the fuel price; IC is the vehicle insurance cost; DR is the vehicle depreciation rate.

8. The vehicle identification method based on a language large model according to claim 7 is characterized in that: The method for determining the recorded vehicle specifically corresponding to the vehicle to be identified is as follows: Based on the vehicle matching index CBO of the vehicle to be identified and the vehicle matching index CBI of each recorded vehicle, calculate the similarity difference d between the vehicle matching index CBO of the vehicle to be identified and the vehicle matching index CBI of each recorded vehicle. The formula used is as follows: ； By comparing all similarity difference d values, find the minimum similarity difference d min , and the corresponding registered vehicle is most similar to the vehicle to be identified, thereby determining the specific registered vehicle corresponding to the vehicle to be identified.

9. The vehicle identification method based on a language large model according to claim 8 is characterized in that: The method for calculating the model adjustment comprehensive evaluation value LPE is as follows: The technical parameter data also includes the number of training rounds AC of the language model, the hidden layer dimension HC of the language model, the number of attention heads EC of the language model, and the word embedding dimension DC of the language model; Based on the number of training rounds AC of the language model, the hidden layer dimension HC of the language model, the number of attention heads EC of the language model, and the word embedding dimension DC of the language model, calculate the model adjustment comprehensive evaluation value LPE. The calculation formula is: ； Among them, COR is the vehicle recognition accuracy rate.

10. The vehicle identification method based on a language large model according to claim 9, characterized in that: The method for judging the model performance level is as follows: The technical parameter data also includes the training duration IM of the language model and the update period FHJ of the language model; Based on the model adjustment comprehensive evaluation value LPE, the training duration IM of the language model, and the update period FHJ of the language model, calculate the model comprehensive performance index MPI. The calculation formula is: ； Among them, PS is the parameter magnitude of the language model; DW is the proportion of vehicle-related information covered in the training data of the language model; The model comprehensive performance index thresholds include the model comprehensive performance index threshold one MPI1 and the model comprehensive performance index threshold two MPI2, and MPI1 < MPI2; When MPI ≤ MPI1, judge that the model performance is poor; When MPI1 < MPI ≤ MPI2, judge that the model performance is medium; When MPI > MPI2, judge that the model performance is good.

Images