Commercial vehicle market segment identification method based on transformer model and contrastive learning

Abstract

The application provides a commercial vehicle market segment identification method based on a Transformer model and contrastive learning, comprising the following steps: S1: obtaining vehicle information and assembling it into a unified sentence; S2: classifying the sentence according to the market segment type to obtain a data set; S3: constructing an input structure based on the data set; S4: constructing a neural network model and training it; and S5: outputting the market segment type to which the vehicle belongs. The method directly extracts vehicle information from a commercial vehicle Internet of Vehicles system and assembles it into a unified sentence, replacing manual telephone collection and eliminating the need for customer service to participate in communication, thus eliminating problems such as the inability to connect the phone and low driver cooperation, improving information collection speed, and saving labor and time costs. The vehicle information is assembled into a unified sentence in a fixed format, avoiding differences in records collected manually. The overall model outputs a probability distribution, and the highest probability is selected as the identification result, avoiding subjective errors in manual judgment.

Description

Technical Field

[0001] This invention relates to the field of vehicle networking technology, and in particular to a method for identifying commercial vehicle market segments based on the Transformer model and contrastive learning. Background Technology

[0002] With the rapid development of IoT, big data, and AI technologies, the connected vehicle industry has also made significant progress. Connected vehicle technology has been widely applied in many fields, including logistics, intelligent transportation, and automobile manufacturing. Connected vehicles can not only provide intelligent diagnostics and remote maintenance for vehicles, but also improve vehicle management and dispatch speed by collecting and analyzing data in real time. Especially in the commercial vehicle sector, the application of connected vehicle technology has enabled big data to gradually demonstrate its value in drivers' production and operation.

[0003] Despite the significant progress made by connected vehicle technology in the commercial vehicle industry, the following drawbacks still exist in terms of segmented market identification:

[0004] 1. Over-reliance on manual labor and low information collection efficiency: For vehicles connected to the freight platform, customer service representatives are dispatched to call and collect information on the drivers' frequently transported goods and routes. During the manual calling process, there are situations where the calls cannot be connected or the drivers have low cooperation in the information collection work, resulting in low information collection efficiency. There are also additional costs such as repeated communication and invalid calls, resulting in excessive costs.

[0005] 2. Low information update frequency and insufficient timeliness: Due to the long update cycle of manually collected information and its susceptibility to changes in the operating environment (such as fluctuations in freight rates and changes in road traffic conditions), the information update frequency is low; the collected data often cannot reflect the latest operating conditions in a timely manner, resulting in insufficient timeliness.

[0006] 3. Low standardization and insufficient accuracy: The information collected manually lacks a unified structured standard. Different customer service representatives have different collection criteria and recording methods, which can easily lead to ambiguous scenarios such as "vague description of goods type" and "unclear definition of route range". The accuracy is insufficient and it is difficult to meet the needs of accurate identification in the commercial vehicle segment market. Summary of the Invention

[0007] The purpose of this invention is to address the shortcomings of existing technologies by proposing a commercial vehicle segmentation identification method based on the Transformer model and contrastive learning.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] A method for identifying commercial vehicle market segments based on Transformer models and contrastive learning includes the following steps:

[0010] S1: Obtain vehicle information and assemble it into a unified sentence structure;

[0011] S11: Filter vehicles that have been tagged with market segment types;

[0012] M vehicles that have been marked with specific market types and verified to be correct are selected from the commercial vehicle Internet of Vehicles system. The specific market types include coal, sand and gravel, green channel, express delivery, grain, chemical, building materials, and others.

[0013] The term "other" refers to market segments other than coal, sand and gravel, green channel, express delivery, grain, chemicals, and building materials.

[0014] The number of vehicles corresponding to each market segment type is denoted as follows: , , , , , , , ;

[0015] S12: Obtain vehicle information and assemble it;

[0016] Obtain vehicle information for M vehicles from the commercial vehicle Internet of Vehicles system. The vehicle information includes vehicle attributes, loading and unloading point information, and vehicle operation information.

[0017] The vehicle attributes include vehicle type, load capacity, etc. The vehicle type includes tractor, cargo truck, dump truck, etc., and the load capacity corresponds to light vehicle, medium vehicle, heavy vehicle, etc.

[0018] The loading and unloading point information includes the latitude and longitude of the loading point and the latitude and longitude of the unloading point;

[0019] The vehicle operation information includes average daily mileage, number of trips per day, and average driving distance;

[0020] The vehicle information corresponding to each vehicle is assembled into a unified sentence according to a set format, resulting in M ​​vehicle information sentences.

[0021] S2: Classify sentence structures according to market segmentation type to obtain a data set;

[0022] The sentence patterns obtained in step S1 are categorized according to the corresponding vehicle market segment type. Each market segment type's sentence pattern forms a data set, resulting in 8 data sets, denoted as . , , , , , , , ;

[0023] Specifically, ={ , , ...... } ;

[0024] ={ , , ...... } ;

[0025] ={ , , ...... } ;

[0026] ={ , , ...... } ;

[0027] ={ , , ...... } ;

[0028] ={ , , ...... } ;

[0029] ={ , , ...... } ;

[0030] ={ , , ...... }

[0031] S3: Construct the input structure based on the dataset;

[0032] For the 8 data sets obtained in step S2, corresponding input structures are formed. Each input structure contains an original sentence (origin), a positive example (entailment), and a negative example (contradiction).

[0033] The original sentence structure is the original sentence structure of each data set; the positive example is the sentence structure in the same data set as the original sentence structure; the negative example is the sentence structure in a different data set than the original sentence structure.

[0034] Let the number of input structures that can be formed by each market segment type be denoted as . ;

[0035] ;

[0036] in, This represents the number of origins for each market segment type, i.e., the total number of elements in the dataset for that market segment type. Greater than or equal to 2; -1 is the number of positive entailments corresponding to each original sentence pattern; M- M represents the number of counterexamples corresponding to each sentence pattern, and M represents the total number of elements in all data sets in step S2.

[0037] 8 Adding them together yields W input structures;

[0038] W= + + + .

[0039] S4: Build and train the neural network model;

[0040] Includes the following sub-steps:

[0041] S41: Convert the sentence structure in the input structure into a token sequence;

[0042] The input structure is converted into a token sequence one by one;

[0043] The original sentence structure origin, positive example entailment, and negative example contradiction in the input structure obtained in step S3 are converted into a fixed-length token sequence one by one using a tokenizer tool (such as Wordpiece, BPE, etc.).

[0044] S42: Build and train a neural network model;

[0045] Build a neural network model using the Encoder part of the Transformer model;

[0046] Set the batch size to V (V is a power of 2, such as 64, 128, etc.), divide the W input structures obtained in step S3 into multiple batches according to the batch size V; use the input structures of each batch to train the neural network model in turn. After each round of training, shuffle the order of the W input structures and re-divide the batches. Repeat the iterative update of the neural network model parameters until the loss value of the neural network model drops to the preset threshold.

[0047] Specifically, the token sequences contained in all batches of input structures in each round are sequentially input into the neural network model, and the corresponding Embedding vectors are output.

[0048] The embedding vector output from the token sequence of the original sentence in each input structure is: The embedding vector output by the token sequence of a positive entailment is: The embedding vector output by the token sequence of the counterexample contradiction is: ;

[0049] use , , The three vectors are compared and learned, and then substituted into the following formula to obtain the loss value for a single batch;

[0050] ;

[0051] Where s is a vector distance calculation function (such as dot product, cosine distance, L1L2 distance, etc.) and e is the natural logarithm;

[0052] The neural network model is initialized with a gradient of 0. Starting from the calculated loss value, the gradient of the output embedding vector and the neural network model parameters is derived in reverse. The neural network model parameters are adjusted by an optimizer such as SGD, combined with the direction of the gradient and the preset learning rate. All batches in this round are iterated. After completion, the samples are shuffled and re-batched. The above process is repeated multiple times until the loss drops to the preset threshold, and the training is completed.

[0053] S5: Outputs the market segment type to which the vehicle belongs;

[0054] S51: Build and train a classifier based on the trained neural network model;

[0055] The following method describes how to build a fully neural network as a classifier and update its network parameters:

[0056] The M vehicle information sentences obtained in step S12 are output as M embedding vectors through a neural network model. The batch size is set to G (G is a power of 2, such as 64, 128, etc.). The M embedding vectors are divided into multiple batches. The classifier network is trained sequentially using each batch. After all batches are trained in each round, the order is shuffled and the batches are re-divided. The classifier model parameters are iterated repeatedly until the loss value drops to a preset threshold.

[0057] The probability of a batch of market segment types is substituted into the following cross-entropy loss function to calculate the difference loss value between the actual market segment type and the predicted market segment type of this batch of data.

[0058] ;

[0059] Where c represents the data number of the Embedding vector in each batch, and the value ranges from 1 to G, where G is a positive integer; This represents the probability distribution of the actual market segmentation of sentence structure. Let i represent the probability distribution of market segment types predicted by the classifier, where i corresponds to 8 market segment types; the probability of the actual market segment type is 1, and the probability of the other 7 market segment types is 0.

[0060] ={ , , , , , , , } ;

[0061] ={ , , , , , , , } ;

[0062] The classifier parameters are adjusted through multiple rounds of iterative training until the differential loss value is within the set threshold, at which point the classifier training is complete.

[0063] The neural network model and the classifier together constitute the overall model;

[0064] S52: Determine the market segment to which the vehicle belongs;

[0065] For unknown vehicles without labeled market segment types, vehicle information is extracted from the commercial vehicle Internet of Vehicles system, assembled into a unified sentence according to the format set in step S12, and input into the overall model. The neural network model in the overall model converts the sentence into the corresponding token sequence output Embedding vector. The classifier performs classification calculation on the Embedding vector and outputs the probability corresponding to the 8 market segment types. The market segment type with the highest probability value is selected as the market segment type to which the unknown vehicle belongs.

[0066] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0067] This method directly extracts vehicle information from the commercial vehicle Internet of Vehicles system and splices it into a unified sentence, replacing manual call collection, eliminating the need for customer service to participate in communication, eliminating problems such as unreachable calls and low driver cooperation, improving information collection speed, avoiding additional consumption such as repeated communication and invalid calls, and saving labor and time costs.

[0068] This method directly connects to the commercial vehicle Internet of Vehicles system. When vehicle information changes, it collects information synchronously to achieve real-time or near-real-time updates, solving the problem of low update frequency of manually collected information and reflecting the latest operating status of the vehicle.

[0069] Multi-dimensional data such as vehicle attributes, loading and unloading points, and operational information are assembled into a unified sentence in a fixed format to avoid discrepancies in manually collected records. A neural network model and a classifier are used to construct an overall model, which outputs the probability distribution of eight market segments. After normalization by the softmax function, the highest probability is selected as the recognition result to avoid subjective errors in human judgment and meet the needs of accurate identification of commercial vehicle market segments. Attached Figure Description

[0070] Figure 1 This is a flowchart illustrating the steps of the commercial vehicle segmentation identification method based on the Transformer model and contrastive learning according to the present invention. Detailed Implementation

[0071] To provide a further understanding of the purpose, structure, features, and functions of the present invention, detailed descriptions are provided below with reference to specific embodiments.

[0072] Example 1

[0073] like Figure 1 As shown, the commercial vehicle market segment identification method based on the Transformer model and contrastive learning includes the following steps:

[0074] S1: Obtain vehicle information and assemble it into a unified sentence structure;

[0075] S11: Filter vehicles that have been tagged with market segment types;

[0076] M vehicles that have been marked with specific market types and verified to be correct are selected from the commercial vehicle Internet of Vehicles system. The specific market types include coal, sand and gravel, green channel, express delivery, grain, chemical, building materials, and others.

[0077] The term "other" refers to market segments other than coal, sand and gravel, green channel, express delivery, grain, chemicals, and building materials.

[0078] The number of vehicles corresponding to each market segment type is denoted as follows:

[0079] , , , , , , , ;

[0080] S12: Obtain vehicle information and assemble it;

[0081] Obtain vehicle information for M vehicles from the commercial vehicle Internet of Vehicles system. The vehicle information includes vehicle attributes, loading and unloading point information, and vehicle operation information.

[0082] The vehicle attributes include vehicle type, load capacity, etc. The vehicle type includes tractor, cargo truck, dump truck, etc., and the load capacity corresponds to light vehicle, medium vehicle, heavy vehicle, etc.

[0083] The loading and unloading point information includes the latitude and longitude of the loading point and the latitude and longitude of the unloading point;

[0084] The vehicle operation information includes average daily mileage, number of trips per day, and average driving distance;

[0085] The vehicle information corresponding to each vehicle is assembled into a unified sentence according to a set format.

[0086] The vehicle type is a tractor, the load capacity is light vehicle, the latitude and longitude of the loading point are (93.50446044569799, 42.8498623825363), the latitude and longitude of the unloading point are (114.52145088724762, 37.85952112941265), the average daily mileage is 800 kilometers, the average number of trips per day is 2, and the average driving distance is 400 kilometers.

[0087] The assembled sentence structure is as follows:

[0088] Vehicle type: Tractor | Load capacity: Light vehicle | Loading points: 93.50446044569799, 42.8498623825363 | Unloading points: 114.52145088724762, 37.85952112941265 | Average daily mileage: 800 km | Average daily trips: 2 | Average driving distance: 400 km;

[0089] Multi-dimensional unstructured vehicle information is automatically converted into structured sentences in a unified format, reducing the need for manual operation from the source.

[0090] S2: Classify sentence structures according to market segmentation type to obtain a data set;

[0091] The sentence patterns obtained in step S1 are categorized according to the corresponding vehicle market segment type. Each market segment type's sentence pattern forms a data set, resulting in 8 data sets, denoted as . , , , , , , , ;

[0092] Specifically, ={ , , ...... };

[0093] ={ , , ...... };

[0094] ={ , , ...... };

[0095] ={ , , ...... };

[0096] ={ , , ...... };

[0097] ={ , , ...... };

[0098] ={ , , ...... };

[0099] ={ , , ...... }

[0100] S3: Construct the input structure based on the dataset;

[0101] For the 8 data sets obtained in step S2, corresponding input structures are formed. Each input structure contains an original sentence (origin), a positive example (entailment), and a negative example (contradiction).

[0102] The original sentence structure is the original sentence structure of each data set; the positive example is the sentence structure in the same data set as the original sentence structure; the negative example is the sentence structure in a different data set than the original sentence structure.

[0103] Let the number of input structures that can be formed by each market segment type be denoted as . ;

[0104] ;

[0105] in, This represents the number of origins for each market segment type, i.e., the total number of elements in the dataset for that market segment type. Greater than or equal to 2; -1 is the number of positive entailments corresponding to each original sentence pattern; M- M represents the number of counterexamples corresponding to each sentence pattern, and M represents the total number of elements in all data sets in step S2.

[0106] 8 Adding them together yields W input structures;

[0107] W= + + + .

[0108] S4: Build and train the neural network model;

[0109] Includes the following sub-steps:

[0110] S41: Convert the sentence structure in the input structure into a token sequence;

[0111] The input structure is converted into a token sequence one by one;

[0112] The original sentence structure origin, positive example entailment, and negative example contradiction in the input structure obtained in step S3 are converted into a fixed-length token sequence one by one using a tokenizer tool (such as Wordpiece, BPE, etc.).

[0113] Specifically, the tokenizer tool decomposes the vehicle information in the original sentence origin, positive example entailment, and negative example contradiction into discrete tokens, maps each token to a unique index value, and forms a fixed-length token sequence, with each sentence corresponding to one token sequence;

[0114] S42: Build and train a neural network model;

[0115] Build a neural network model using the Encoder part of the Transformer model;

[0116] Set the batch size to V (V is a power of 2, such as 64, 128, etc.), divide the W input structures obtained in step S3 into multiple batches according to the batch size V; use the input structures of each batch to train the neural network model in turn. After each round of training, shuffle the order of the W input structures and re-divide the batches. Repeat the iterative update of the neural network model parameters until the loss value of the neural network model drops to the preset threshold.

[0117] Specifically, the token sequences contained in all batches of input structures in each round are sequentially input into the neural network model, and the corresponding Embedding vectors are output.

[0118] The embedding vector output from the token sequence of the original sentence in each input structure is: The embedding vector output by the token sequence of a positive entailment is: The embedding vector output by the token sequence of the counterexample contradiction is: ;

[0119] use , , The three vectors are compared and learned, and then substituted into the following formula to obtain the loss value for a single batch;

[0120] ;

[0121] Where s is a vector distance calculation function (such as dot product, cosine distance, L1L2 distance, etc.) and e is the natural logarithm;

[0122] The neural network model is initialized with a gradient of 0. Starting from the calculated loss value, the gradient of the output embedding vector and the neural network model parameters is derived in reverse. The neural network model parameters are adjusted by an optimizer such as SGD, combined with the direction of the gradient and the preset learning rate. All batches in this round are iterated. After completion, the samples are shuffled and re-batched. The above process is repeated multiple times until the loss drops to the preset threshold, and the training is completed.

[0123] The optimizer adjusts the values ​​of each parameter based on the direction of the gradient (a positive gradient indicates that the parameter needs to be decreased, and a negative gradient indicates that the parameter needs to be increased) and the preset learning rate.

[0124] S5: Outputs the market segment type to which the vehicle belongs;

[0125] S51: Build and train a classifier based on the trained neural network model;

[0126] The following method describes how to build a fully neural network as a classifier and update its network parameters:

[0127] The M vehicle information sentences obtained in step S12 are output as M embedding vectors through a neural network model. The batch size is set to G (G is a power of 2, such as 64, 128, etc.). The M embedding vectors are divided into multiple batches. The classifier network is trained sequentially using each batch. After all batches are trained in each round, the order is shuffled and the batches are re-divided. The classifier model parameters are iterated repeatedly until the loss value drops to a preset threshold.

[0128] The probability of a batch of market segment types is substituted into the following cross-entropy loss function to calculate the difference loss value between the actual market segment type and the predicted market segment type of this batch of data.

[0129] ;

[0130] Where c represents the data number of the Embedding vector in each batch, and the value ranges from 1 to G, where G is a positive integer; This represents the probability distribution of the actual market segmentation of sentence structure. Let i represent the probability distribution of market segment types predicted by the classifier, where i corresponds to 8 market segment types; the probability of the actual market segment type is 1, and the probability of the other 7 market segment types is 0.

[0131] ={ , , , , , , , };

[0132] ={ , , , , , , , };

[0133] The classifier parameters are adjusted through multiple rounds of iterative training until the differential loss value is within the set threshold, at which point the classifier training is complete.

[0134] The neural network model and the classifier together constitute the overall model;

[0135] Outputting the probability distribution of each category not only helps determine the market segment type, but also allows us to judge the reliability of the identification results through probability values, facilitating subsequent verification and optimization.

[0136] S52: Determine the market segment to which the vehicle belongs;

[0137] For unknown vehicles without labeled market segment types, vehicle information is extracted from the commercial vehicle Internet of Vehicles system, assembled into a unified sentence according to the format set in step S12, and input into the overall model. The neural network model in the overall model converts the sentence into the corresponding token sequence output Embedding vector. The classifier performs classification calculation on the Embedding vector and outputs the probability corresponding to the 8 market segment types. The market segment type with the highest probability value is selected as the market segment type to which the unknown vehicle belongs.

[0138] The present invention has been described in the above-described embodiments; however, these embodiments are merely examples for implementing the present invention. It must be noted that the disclosed embodiments do not limit the scope of the present invention. Conversely, any modifications and refinements made without departing from the spirit and scope of the present invention are within the scope of patent protection of the present invention.

Claims

1. A method for identifying commercial vehicle market segments based on Transformer models and contrastive learning, characterized in that: Includes the following steps: S1: Obtain vehicle information and assemble it into a unified sentence structure; S11: Filter vehicles that have been tagged with market segment types; S12: Obtain vehicle information and assemble it; S2: Classify sentence structures according to market segmentation type to obtain a data set; S3: Construct the input structure based on the dataset; For the data set obtained in step S2, a corresponding input structure is formed. Each input structure contains an original sentence (origin), a positive example (entailment), and a negative example (contradiction). The original sentence structure is the original sentence structure of each data set; the positive example is the sentence structure in the same data set as the original sentence structure. The counterexample is a sentence that is not in the same data set as the original sentence; Let the number of input structures that can be formed by each market segment type be denoted as . ; ； in, This represents the number of origins for each market segment type, i.e., the total number of elements in the dataset for that market segment type. Greater than or equal to 2; -1 is the number of positive entailments corresponding to each original sentence pattern; M- M represents the number of counterexamples corresponding to each sentence pattern, and M represents the total number of elements in all data sets in step S2. 8 Adding them together yields W input structures; S4: Build and train the neural network model; S41: Convert the sentence structure in the input structure into a token sequence; The input structure is converted into a token sequence one by one; The tokenizer tool transforms the original sentence structure origin, positive example entailment, and negative example contradiction in the input structure obtained in step S3 into a fixed-length token sequence one by one. S42: Build and train a neural network model; Build a neural network model using the Encoder part of the Transformer model; Set the batch size to V, and divide the W input structures obtained in step S3 into multiple batches according to the batch size V; use the input structures of each batch to train the neural network model in turn. After each round of training, shuffle the order of the W input structures and re-divide the batches. Repeat the iterative update of the neural network model parameters until the loss value of the neural network model drops to the preset threshold. The token sequence contained in all batches of input structure in each round is sequentially input into the neural network model, and the corresponding embedding vector is output. The embedding vector output from the token sequence of the original sentence in each input structure is: The embedding vector output by the token sequence of a positive entailment is: The embedding vector output by the token sequence of the counterexample contradiction is: ; use , , The three vectors are compared and learned, and then substituted into the following formula to obtain the loss value for a single batch; ； Where s is the vector distance calculation function and e is the natural logarithm; The neural network model is initialized with a gradient of 0. Starting from the calculated loss value, the gradient of the output embedding vector and the neural network model parameters is derived in reverse. The neural network model parameters are adjusted by the SGD optimizer, taking into account the direction of the gradient and the preset learning rate. The process iterates through all batches in this round. After completion, the samples are shuffled and re-batched. The above process is repeated multiple times until the loss drops to the preset threshold, and the training is completed. S5: Outputs the market segment type to which the vehicle belongs; S51: Build and train a classifier based on the trained neural network model; S52: Determine the market segment to which the vehicle belongs.

2. The commercial vehicle market segmentation identification method based on Transformer model and contrastive learning as described in claim 1, characterized in that: The specific details of step S1 are as follows: S11: Filter vehicles that have been tagged with market segment types; M vehicles that have been marked with specific market types and verified to be correct are selected from the commercial vehicle Internet of Vehicles system. The specific market types include coal, sand and gravel, green channel, express delivery, grain, chemical, building materials, and others. The term "other" refers to market segments other than coal, sand and gravel, green channel, express delivery, grain, chemicals, and building materials. The number of vehicles corresponding to each market segment type is denoted as follows: 、 、 、 、 、 、 、 ； S12: Obtain vehicle information and assemble it; Obtain vehicle information for M vehicles from the commercial vehicle Internet of Vehicles system. The vehicle information includes vehicle attributes, loading and unloading point information, and vehicle operation information. The vehicle information corresponding to each vehicle is assembled into a unified sentence according to a set format, resulting in M ​​vehicle information sentences.

3. The commercial vehicle segmentation identification method based on Transformer model and contrastive learning as described in claim 1, characterized in that: The specific details of step S2 are as follows: The sentence patterns obtained in step S1 are categorized according to the corresponding vehicle market segment type. Each market segment type's sentence pattern forms a data set, resulting in 8 data sets, denoted as . , , , , , , , .

4. The commercial vehicle market segmentation identification method based on Transformer model and contrastive learning as described in claim 1, characterized in that: The specific details of step S5 are as follows: S51: Build and train a classifier based on the trained neural network model; The following method describes how to build a fully neural network as a classifier and update its network parameters: The M vehicle information sentences obtained in step S12 are output as M embedding vectors through a neural network model. The batch size is set to G, and the M embedding vectors are divided into multiple batches. The classifier network is trained sequentially using each batch. After all batches are trained in each round, the order is shuffled and the batches are re-divided. The classifier model parameters are iterated repeatedly until the loss value drops to a preset threshold. The probability of a batch of market segment types is substituted into the following cross-entropy loss function to calculate the difference loss value between the actual market segment type and the predicted market segment type of this batch of data. ； Where c represents the data number of the Embedding vector in each batch, and the value ranges from 1 to G, where G is a positive integer; This represents the probability distribution of the actual market segmentation of sentence structure. Let i represent the probability distribution of market segment types predicted by the classifier, where i corresponds to 8 market segment types; the probability of the actual market segment type is 1, and the probability of the other 7 market segment types is 0. ={ ， ， ， ， ， ， ， } ； ={ ， ， ， ， ， ， ， } ； The classifier parameters are adjusted through multiple rounds of iterative training until the differential loss value is within the set threshold, at which point the classifier training is complete. The neural network model and the classifier together constitute the overall model; S52: Determine the market segment to which the vehicle belongs; For unknown vehicles without labeled market segment types, vehicle information is extracted from the commercial vehicle Internet of Vehicles system, assembled into a unified sentence according to the format set in step S12, and input into the overall model. The neural network model in the overall model converts the sentence into the corresponding token sequence output Embedding vector. The classifier performs classification calculation on the Embedding vector and outputs the probability corresponding to the 8 market segment types. The market segment type with the highest probability value is selected as the market segment type to which the unknown vehicle belongs.

Images