Commercial truck portrait construction and label assignment method based on multi-source heterogeneous data fusion

Abstract

The present invention discloses a commercial truck portrait construction and label assignment method under multi-source heterogeneous data fusion, which belongs to the field of data processing technology, and specifically includes: by collecting multi-source heterogeneous data, using adaptive preprocessing and principal component analysis to extract key features, and constructing high-level features for fusion, to form a multi-dimensional feature data set; based on the data set, using a composite reinforcement learning algorithm in combination with multimodal data to construct a commercial truck portrait, and assign accurate portrait labels; the system introduces a causal inference method to detect deviations and anomalies in label assignment, and establishes a feedback mechanism to optimize machine learning model parameters and portrait system; truck portraits are applied in actual business, feedback is collected for continuous optimization, and a personalized service recommendation system is developed to improve the intelligence level of the logistics industry and enhance service personalization and accuracy.

Description

Technical Field

[0001] The present invention belongs to the field of data processing technology, and specifically to a commercial truck portrait construction and label assignment method under multi-source heterogeneous data fusion. Background Art

[0002] Commercial truck profiling is a process of integrating the multi-dimensional feature data of commercial trucks into a structured form that is easy to understand and use, including the basic attributes of the vehicle: such as brand, model, tonnage, operating status: such as mileage, maintenance records, driving behavior: such as frequently traveled routes, average speed information. Through data mining and machine learning technology, these complex data are converted into a series of labels or feature vectors to form a unique portrait of each vehicle.

[0003] Label assignment is the process of assigning one or more category labels to each truck based on the construction of truck portraits and according to the specific attributes and behavior patterns of the vehicle. These labels can be vehicle type: such as refrigerated trucks, flatbed trucks, usage frequency: such as high-frequency use, low-frequency standby, maintenance status: such as good, needing repair, and label assignment aims to quickly identify and distinguish trucks with different characteristics through concise and clear labels to support efficient decision-making and automated processes.

[0004] For example, the Chinese patent application with publication number CN111291236A discloses a method and system for obtaining and analyzing user portrait information of a highway network, including: using user portrait technology, establishing an ETC user portrait construction solution model and developing a set of ETC user portrait analysis prototype systems, constructing user basic attributes, user consumption, user travel time and space habits, user value, user activity, user credit and other labels for highway ETC users, and describing and characterizing users from multiple angles. Based on user grouping, the travel habits of different users are characterized, and the correlation, differences and reasons behind passenger and truck travel are further analyzed; from a time perspective, the impact of holidays, day and night on user travel is analyzed. The system is based on the proposed ETC user portrait model, and is designed based on SQL database and python language. This technical solution has the advantages of high efficiency, low computing cost and strong real-time results, and can meet actual application needs.

[0005] The above existing technologies all have the following problems: 1) The precision is low when faced with complex and changeable traffic environments and user behaviors; 2) When faced with rapidly changing traffic conditions and user needs, their dynamic adjustment capabilities may be limited; 3) The application scenarios are relatively single. Summary of the invention

[0006] In view of the shortcomings of the prior art, the present invention proposes a commercial truck portrait construction and label assignment method under multi-source heterogeneous data fusion. By collecting multi-source heterogeneous data, adaptive preprocessing and principal component analysis are used to extract key features, and high-level features are constructed for fusion to form a multi-dimensional feature data set; based on the data set, a composite reinforcement learning algorithm is used in combination with multimodal data to construct commercial truck portraits, and accurate portrait labels are assigned; the system introduces causal inference methods to detect deviations and anomalies in label assignment, and establishes a feedback mechanism to optimize machine learning model parameters and portrait systems; truck portraits are applied in actual business, feedback is collected for continuous optimization, and a personalized service recommendation system is developed to improve the intelligence level of the logistics industry and enhance service personalization and accuracy.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] The commercial truck portrait construction and label assignment method based on multi-source heterogeneous data fusion includes:

[0009] Step S1: Collect multi-source heterogeneous data, and use an adaptive algorithm to preprocess them. Use the principal component analysis method to extract features from the preprocessed multi-source heterogeneous data. At the same time, based on the preprocessed multi-source heterogeneous data, construct high-level features through a feature construction method, and fuse the feature extraction results with the high-level features to obtain a multi-dimensional feature data set.

[0010] Step S2: Based on the multi-dimensional feature data set, a portrait construction strategy is used to construct a commercial truck portrait, multimodal data is introduced, and a composite reinforcement learning algorithm is used to assign corresponding portrait labels to the truck;

[0011] Step S3: Use causal inference methods to identify deviations and anomalies in label assignment, establish a feedback mechanism, feed back the detected deviations and anomalies to the profile building system, adjust the parameters of the machine learning model, and iteratively optimize the profile building system and label assignment strategy based on the feedback results;

[0012] Step S4: Apply truck portraits to actual business scenarios, collect feedback on application effects, continuously iterate and optimize the portrait construction system, and at the same time, use truck portrait data and combine user behavior analysis to develop a personalized service recommendation system;

[0013] The specific steps of step S2 include:

[0014] S2.1: Obtaining a multi-dimensional feature dataset ,in, Indicates the basic information characteristics of the truck. Represents the characteristics of truck operation data, Represents the characteristics of environmental data, represents the behavior data characteristics of truck drivers, n represents the number of basic information characteristics of trucks, m represents the number of operation data characteristics of trucks, p represents the number of environmental data characteristics, and q represents the number of behavior data characteristics of truck drivers;

[0015] S2.2: Define the portrait dimension according to the category of the multi-dimensional feature dataset. According to the portrait dimension, use the label calculation method to divide the label into fact label, model label and advanced label system. Based on the label calculation method and the multi-dimensional feature dataset X, set the value range for each label.

[0016] The specific steps of step S2 also include:

[0017] S2.3: Perform cluster analysis on X. After clustering, use the classification strategy to divide each type of feature, and use deep reinforcement learning to combine the clustering-classification results to assign labels to each type of feature;

[0018] S2.4: Based on the label system, build a truck portrait and use evaluation indicators to evaluate the portrait construction effect. According to the evaluation results, adjust the feature selection and composite reinforcement learning algorithm parameters to optimize the portrait construction process;

[0019] The specific steps of S2.3 include:

[0020] S2.31: Set the number of clusters and distance measurement method according to the silhouette coefficient method, use the clustering strategy to divide X into clusters, calculate the silhouette coefficient of each data point, and use the improved silhouette algorithm to evaluate the clustering effect. The formula is:

[0021]

[0022] in, represents the comprehensive score of commercial truck i, represents the weight of the jth indicator, and represents the lowest and highest performance index values ​​of truck i under the jth index, and Indicates that it is used to adjust and The weight coefficient in the calculation is, represents the offset, j represents different indicator types, and r represents the total number of indicator types;

[0023] S2.32: Based on the clustering results, analyze the characteristics of each cluster, formulate a classification strategy based on different characteristics, and use the classification strategy to assign labels to each cluster after clustering;

[0024] The specific steps of S2.3 also include:

[0025] S2.33: Use deep reinforcement learning algorithm to learn the optimal behavior strategy based on the characteristics of the truck in different states, where the cluster label is used as the state space;

[0026] S2.34: Based on the actual application results, collect feedback information, iteratively optimize the clustering algorithm, classification strategy, and deep reinforcement learning model, and adjust the parameters and structure of the deep reinforcement learning model.

[0027] Specifically, the label calculation method in S2.2 specifically includes:

[0028] S2.21: Determine the label calculation method and value;

[0029] If it is a fact label, it is assigned a value based on the feature value in the data set and the value range is set [ a , b ] , where a and b represent the minimum and maximum values ​​of the fact label value range respectively;

[0030] If it is an advanced label, it is calculated based on multiple fact labels through weighted summation and the value range is set. [ c , d ] , where c and d represent the minimum and maximum values ​​of the advanced tag value range, respectively;

[0031] If it is a model label, the label value is calculated using regression statistics, and the level is set to high, medium, or low according to the label value;

[0032] S2.22: Use the defined label calculation method and X to calculate the label value of each feature object, and assign the calculated label value to the corresponding feature object to build a label system.

[0033] Specifically, the specific formula of the deep reinforcement learning algorithm in S2.33 is:

[0034]

[0035] in, Representation parameters The gradient of represents the gradient function, Indicated in strategy The expectations below, represents the action at time t, Represents the policy function, which is in the state Next select action The probability of represents the state at time t, represents the advantage function, which is the cumulative reward from time t to the terminal state, represents the discount factor, Indicates status The value function under Indicates status The value function under represents the baseline function, and T represents the time length.

[0036] An electronic device includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of a commercial truck portrait construction and label assignment method under multi-source heterogeneous data fusion are implemented.

[0037] A computer-readable storage medium stores computer instructions, which, when executed, execute the steps of a commercial truck portrait construction and label assignment method under multi-source heterogeneous data fusion.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] 1. The present invention proposes a commercial truck portrait construction and label assignment method under multi-source heterogeneous data fusion, which improves the comprehensiveness and diversity of data by collecting and fusing data from different sources, such as sensors, social media, satellite images, mobile applications, and different types, such as structured, semi-structured, and unstructured data; an adaptive algorithm is used to preprocess the data, including data cleaning, data integration, and data conversion steps, which effectively removes noise and outliers in the data and improves the quality and availability of the data; by performing principal component analysis on the preprocessed multi-source heterogeneous data, the main features in the data are extracted, the dimension of the data is reduced, and key information is retained; advanced features are constructed by a feature construction method, and the feature extraction results are fused with the advanced features to obtain a multi-dimensional feature data set. This fusion method further enriches the feature dimensions of the portrait and improves the accuracy and richness of the portrait construction.

[0040] 2. The present invention proposes a commercial truck portrait construction and label assignment method under multi-source heterogeneous data fusion. Based on a multi-dimensional feature data set, a portrait construction strategy is adopted to construct a commercial truck portrait, and multimodal data and a composite reinforcement learning algorithm are introduced to assign corresponding portrait labels to the truck. This method can dynamically adjust the label assignment strategy according to the actual operating conditions and market environment of the truck, thereby improving the accuracy and applicability of the label; it uses a causal inference method to identify deviations and anomalies in label assignment, and establishes a feedback mechanism, which can promptly discover and correct errors and unreasonable aspects in label assignment to ensure the accuracy and reliability of the portrait. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a schematic diagram of the commercial truck portrait construction and label assignment method under multi-source heterogeneous data fusion of the present invention;

[0042] Figure 2 This is a principle flow chart of the commercial truck portrait construction and label assignment method under multi-source heterogeneous data fusion of the present invention;

[0043] Figure 3 It is a label allocation flow chart of the commercial truck portrait construction and label allocation method under multi-source heterogeneous data fusion of the present invention;

[0044] Figure 4 This is a flow chart for implementing the clustering strategy of the commercial truck portrait construction and label assignment method under multi-source heterogeneous data fusion of the present invention. DETAILED DESCRIPTION

[0045] Example 1

[0046] See also Figure 1-Figure 2 , an embodiment provided by the present invention: a commercial truck portrait construction and label assignment method under multi-source heterogeneous data fusion, comprising the following steps:

[0047] Step S1: Collect multi-source heterogeneous data, and use an adaptive algorithm to preprocess them, extract features from the preprocessed multi-source heterogeneous data, and construct high-level features based on the preprocessed multi-source heterogeneous data through a feature construction method, and fuse the feature extraction results with the high-level features to obtain a multi-dimensional feature data set;

[0048] Among them, multi-source heterogeneous data include: basic vehicle information, such as model, brand, and load; driving data, such as mileage, average speed, and fuel consumption; maintenance records, such as the number of maintenance times and maintenance types.

[0049] Furthermore, the specific steps of step S1 include:

[0050] (1) Data collection and organization: determine the data source, collect data, and conduct preliminary organization and classification;

[0051] (2) Data preprocessing: Perform data cleaning, standardization and normalization according to the data preprocessing steps, and apply adaptive algorithms to optimize data quality;

[0052] (3) Feature extraction and selection: Extract basic features from the preprocessed data and remove redundant and noisy features through feature selection methods;

[0053] (4) Advanced feature construction: construct advanced features based on preprocessed data, and fuse basic features with advanced features to form a multi-dimensional feature data set;

[0054] (5) Result evaluation and application: Evaluate the fused feature data set.

[0055] Step S2: Based on the multi-dimensional feature data set, a portrait construction strategy is used to construct a commercial truck portrait, multimodal data is introduced, and a composite reinforcement learning algorithm is used to assign corresponding portrait labels to the truck;

[0056] Step S3: Use causal inference methods to identify deviations and anomalies in label assignment, establish a feedback mechanism, feed back the detected deviations and anomalies to the profile building system, adjust the parameters of the machine learning model, and iteratively optimize the profile building system and label assignment strategy based on the feedback results;

[0057] Furthermore, the specific steps of step S3 include:

[0058] (1) Identify goals and problems: Identify the types of deviations and anomalies in label assignments to be identified, and set specific goals, such as improving the accuracy of label assignments and reducing misassignments;

[0059] (2) Data collection and preprocessing: Collect data related to label assignment, including user behavior data and portrait feature data, and clean and preprocess the data, such as removing noise, processing missing values, and standardizing;

[0060] (3) Construct causal graphs and identify biases: Use causal inference methods, such as Do-calculus, to construct causal graphs and identify confounding variables and biases that affect label assignment in the causal graphs;

[0061] (4) Design anomaly detection algorithm: Select statistical anomaly detection based on data characteristics and problem requirements;

[0062] (5) Apply anomaly detection algorithms to monitor the label assignment process and identify potential deviations and anomalies;

[0063] (6) Establish a feedback mechanism: Design feedback channels, such as log systems and monitoring dashboards, to collect anomaly detection results and set up feedback processes to ensure that detected deviations and anomaly information can be fed back to the profile building system in a timely and accurate manner;

[0064] (7) Adjusting machine learning model parameters: Based on the feedback deviation and abnormal information, analyze their impact on the machine learning model and adjust the model parameters, such as learning rate and regularization coefficient, to optimize the model performance;

[0065] (8) Iteratively optimize the portrait system and label allocation strategy: Re-evaluate the effectiveness of the portrait system and label allocation strategy based on the adjusted model parameters and feedback results, and continuously iterate and optimize until satisfactory results are achieved.

[0066] Step S4: Apply truck portraits to actual business scenarios, collect feedback on application effects, and continuously iterate and optimize the portrait construction system. At the same time, use truck portrait data and combine it with user behavior analysis to develop a personalized service recommendation system.

[0067] Among them, the recommendation system in the present invention adopts a collaborative filtering recommendation algorithm, and the collaborative filtering recommendation algorithm is the existing technology content in this field, is not an inventive solution of the present application, and will not be described here.

[0068] Example 2

[0069] See also Figure 3-Figure 4 In this embodiment, the specific steps of step S2 include:

[0070] S2.1: Obtaining a multi-dimensional feature dataset ,in, Indicates the basic information characteristics of the truck. Represents the characteristics of truck operation data, Represents the characteristics of environmental data, represents the behavior data characteristics of truck drivers, n represents the number of basic information characteristics of trucks, m represents the number of operation data characteristics of trucks, p represents the number of environmental data characteristics, and q represents the number of behavior data characteristics of truck drivers;

[0071] Among them, the basic information characteristics of trucks include: brand code, model code, load capacity, and vehicle size; the operating data characteristics of trucks include: driving trajectory characteristics, transportation efficiency characteristics, and fuel consumption characteristics; the environmental data characteristics include: weather characteristics and road conditions characteristics; the behavioral data characteristics of truck drivers include: driving habit characteristics and rest time characteristics.

[0072] S2.2: Define the portrait dimension according to the category of the multi-dimensional feature dataset. According to the portrait dimension, use the label calculation method to divide the label into fact label, model label and advanced label system. Based on the label calculation method and the multi-dimensional feature dataset X, set the value range for each label.

[0073] Furthermore, the specific steps of S2.2 include:

[0074] (1) Analyze the categories and contents of the multi-dimensional feature dataset, identify the key dimensions that can describe the characteristics of trucks, and define the specific meaning and scope of each portrait dimension;

[0075] (2) The labels are divided into fact labels, model labels and advanced label systems, where:

[0076] a. Factual labels: directly derived from labels that already exist in the dataset and can be obtained without calculation, such as car model and brand;

[0077] b. Model labels: labels that need to be calculated through statistical models or machine learning algorithms, such as predicting failure rates and evaluating transportation efficiency;

[0078] c. Advanced labeling system: A higher-level label obtained by further combining or analyzing model labels and fact labels, such as comprehensive scores and risk levels;

[0079] (3) For each tag, set a value range based on its type and business requirements. The value range of fact tags is usually fixed, such as a list of strings for car models and a set of specific brands for brands.

[0080] (4) Design computational methods for each model label and advanced label system using machine learning model prediction methods;

[0081] (5) Using the defined label calculation method and the multi-dimensional feature dataset X, calculate all label values ​​for each feature object and ensure the accuracy and efficiency of label calculation;

[0082] (6) Organize the calculated label values ​​into a structured label system and verify the label system using business data or simulation data to ensure the accuracy and effectiveness of the labels.

[0083] S2.3: Perform cluster analysis on X. After clustering, use the classification strategy to divide each type of feature, and use deep reinforcement learning to combine the clustering-classification results to assign labels to each type of feature;

[0084] S2.4: Based on the label system, build a truck portrait and use evaluation indicators to evaluate the portrait construction effect. According to the evaluation results, adjust the feature selection and composite reinforcement learning algorithm parameters to optimize the portrait construction process.

[0085] The label calculation method in S2.2 specifically includes:

[0086] S2.21: Determine the label calculation method and value;

[0087] If it is a fact label, it is assigned a value based on the feature value in the data set and the value range is set [ a , b ] , where a and b represent the minimum and maximum values ​​of the fact label value range respectively;

[0088] If it is an advanced label, it is calculated based on multiple fact labels through weighted summation and the value range is set. [ c , d ] , where c and d represent the minimum and maximum values ​​of the advanced tag value range, respectively;

[0089] If it is a model label, the label value is calculated using regression statistics, and the level is set to high, medium, or low according to the label value;

[0090] S2.22: Use the defined label calculation method and X to calculate the label value of each feature object, and assign the calculated label value to the corresponding feature object to build a label system.

[0091] Furthermore, the specific steps of S2.22 include:

[0092] (1) Receive multi-dimensional feature data sets and pre-process them, including cleaning, conversion, and standardization, to ensure data quality and consistency;

[0093] (2) Clarify the type of label that needs to be calculated and its business meaning, and use statistical algorithms to design the label calculation method. For example, for the "efficient transportation" label, the calculation method is designed to be based on the ratio of mileage and fuel consumption; for the "high risk" label, the calculation method is designed to be based on maintenance records and accident history.

[0094] (3) Calculate the label value of each feature object: Apply the defined label calculation method to calculate each feature object in the multi-dimensional feature data set, where the feature object refers to the truck;

[0095] (4) According to the calculation results, a label value is assigned to each feature object, where the label value is in a composite form and contains a vector of multiple dimensions;

[0096] (5) Organize the calculated tag values ​​into a structured tag system to ensure that there is a clear hierarchy and relationship between tags;

[0097] (6) Design interfaces for tag query, retrieval, and management to facilitate subsequent application and maintenance.

[0098] The specific steps of S2.3 include:

[0099] S2.31: Set the number of clusters and distance measurement method according to the silhouette coefficient method, use the clustering strategy to divide X into clusters, calculate the silhouette coefficient of each data point, and use the improved silhouette algorithm to evaluate the clustering effect. The formula is:

[0100]

[0101] in, represents the comprehensive score of commercial truck i, represents the weight of the jth indicator, and represents the lowest and highest performance index values ​​of truck i under the jth index, and Indicates that it is used to adjust and The weight coefficient in the calculation is, represents the offset, j represents different indicator types, and r represents the total number of indicator types;

[0102] S2.32: Based on the clustering results, analyze the characteristics of each cluster, formulate a classification strategy based on different characteristics, and use the classification strategy to assign labels to each cluster after clustering;

[0103] Furthermore, the specific steps of S2.32 include:

[0104] (1) Cluster analysis: Select K-means clustering to perform cluster analysis on data set X. According to the characteristics of the clustering algorithm and the characteristics of the data set, set parameters such as the number of clusters and the size of the neighborhood, and execute the clustering algorithm to obtain the clustering results, that is, divide the data set X into multiple clusters.

[0105] (2) Analyze the characteristics of each cluster: Calculate the statistical characteristics of each cluster, such as the mean, standard deviation, median, and mode. Analyze the distribution characteristics of the data points in the cluster, such as density and shape, based on the statistical characteristics. Identify the characteristics of each cluster. These characteristics can be numerical or categorical.

[0106] (3) Formulate a classification strategy: Develop a corresponding classification strategy based on the characteristics of each cluster. The classification strategy can be based on the size, range, and distribution of the feature values, or it can be developed in combination with business knowledge and domain knowledge.

[0107] (4) Assign a label to each cluster after clustering: According to the established classification strategy, use an automated algorithm to assign one or more labels to each cluster. The labels can clearly represent the characteristics and category information of the cluster.

[0108] S2.33: Use deep reinforcement learning algorithm to learn the optimal behavior strategy based on the characteristics of the truck in different states, where the cluster label is used as the state space;

[0109] Furthermore, the specific steps of S2.33 include:

[0110] (1) Data collection and preprocessing: Collect truck operation data, clean and preprocess it, and use clustering algorithms to divide the state space;

[0111] (2) Model construction: Select and build a deep reinforcement learning model, define the state space, action space, and reward function;

[0112] (3) Initialization: Initialize the neural network parameters and set hyperparameters such as learning rate and discount factor;

[0113] (4) Training loop:

[0114] 1) Get the current state (cluster label) from the environment;

[0115] 2) Select actions based on current state and strategy;

[0116] 3) Execute the action and obtain the next state and reward;

[0117] 4) Store the experience (state, action, reward, next state) in the experience replay buffer;

[0118] 5) Randomly extract a batch of samples from the experience replay buffer, calculate the target value, and update the neural network parameters.

[0119] (5) Testing and evaluation: Evaluate model performance on the test set and perform fine-tuning as needed;

[0120] (6) Deployment and application: Deploy the trained model to the actual system for real-time decision-making.

[0121] S2.34: Based on the actual application results, collect feedback information, iteratively optimize the clustering algorithm, classification strategy, and deep reinforcement learning model, and adjust the parameters and structure of the deep reinforcement learning model.

[0122] The specific formula of the deep reinforcement learning algorithm in S2.33 is:

[0123]

[0124] in, Representation parameters The gradient of Represents the gradient function, which measures the policy At the given parameters Good or bad, Indicated in strategy The expectation under is a mapping from state to action. represents the action at time t, Represents the policy function, which is in the state Next select action The probability of represents the state at time t, represents the advantage function, which is the cumulative reward from time t to the terminal state, represents the discount factor used to calculate the present value of future rewards, and λ ∈ [ 0 , 1 ] , which is used to ensure that the sum of rewards for infinitely long trajectories is finite, if , the more the agent values ​​immediate rewards; , the more the agent values ​​future rewards, Indicates status The value function under Indicates status The value function under Represents the baseline function, which is used to reduce the variance of the gradient estimate without changing the direction of the gradient. It is usually independent of the action and only related to the state. At the same time, the introduction of the baseline function can speed up the learning speed and improve the stability of learning. T represents the time length.

[0125] Example 3

[0126] An electronic device includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of a method for constructing a commercial truck portrait and assigning labels under multi-source heterogeneous data fusion are implemented. For details, please refer to the above method embodiment, which will not be repeated here.

[0127] A computer-readable storage medium stores computer instructions, which, when executed, execute the steps of a commercial truck portrait construction and label assignment method under multi-source heterogeneous data fusion, wherein the storage medium may be a volatile or non-volatile computer-readable storage medium.

[0128] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation modes, which are merely illustrative rather than restrictive. Under the guidance of the present invention, ordinary technicians in the field may also change, modify, replace and modify the above-mentioned embodiments without departing from the purpose and scope of protection of the present invention, and all of these are within the protection of the present invention.

Claims

1. A commercial truck image construction and label assignment method based on multi-source heterogeneous data fusion, characterized in that: include: Step S1: Collect multi-source heterogeneous data, and use an adaptive algorithm to preprocess them. Use the principal component analysis method to extract features from the preprocessed multi-source heterogeneous data. At the same time, based on the preprocessed multi-source heterogeneous data, construct high-level features through a feature construction method, and fuse the feature extraction results with the high-level features to obtain a multi-dimensional feature data set. Step S2: Based on the multi-dimensional feature data set, a portrait construction strategy is used to construct a commercial truck portrait, multimodal data is introduced, and a composite reinforcement learning algorithm is used to assign corresponding portrait labels to the truck; Step S3: Use causal inference methods to identify deviations and anomalies in label assignment, establish a feedback mechanism, feed back the detected deviations and anomalies to the profile building system, adjust the parameters of the machine learning model, and iteratively optimize the profile building system and label assignment strategy based on the feedback results; Step S4: Apply truck portraits to actual business scenarios, collect feedback on application effects, continuously iterate and optimize the portrait construction system, and at the same time, use truck portrait data and combine user behavior analysis to develop a personalized service recommendation system; The specific steps of step S2 include: S2.1: Obtaining a multi-dimensional feature dataset ,in, Indicates the basic information characteristics of the truck. Represents the characteristics of truck operation data, Represents the characteristics of environmental data, represents the behavior data characteristics of truck drivers, n represents the number of basic information characteristics of trucks, m represents the number of operation data characteristics of trucks, p represents the number of environmental data characteristics, and q represents the number of behavior data characteristics of truck drivers; S2.2: Define the portrait dimension according to the category of the multi-dimensional feature dataset. According to the portrait dimension, use the label calculation method to divide the label into fact label, model label and advanced label system. Based on the label calculation method and the multi-dimensional feature dataset X, set the value range for each label. The specific steps of step S2 also include: S2.3: Perform cluster analysis on X. After clustering, use the classification strategy to divide each type of feature, and use deep reinforcement learning to combine the clustering-classification results to assign labels to each type of feature; S2.4: Based on the label system, build a truck portrait and use evaluation indicators to evaluate the portrait construction effect. According to the evaluation results, adjust the feature selection and composite reinforcement learning algorithm parameters to optimize the portrait construction process; The specific steps of S2.3 include: S2.31: Set the number of clusters and distance measurement method according to the silhouette coefficient method, use the clustering strategy to divide X into clusters, calculate the silhouette coefficient of each data point, and use the improved silhouette algorithm to evaluate the clustering effect. The formula is: ； in, represents the comprehensive score of commercial truck i, represents the weight of the jth indicator, and represents the lowest and highest performance index values ​​of truck i under the jth index, and Indicates that it is used to adjust and The weight coefficient in the calculation is, represents the offset, j represents different indicator types, and r represents the total number of indicator types; S2.32: Based on the clustering results, analyze the characteristics of each cluster, formulate a classification strategy based on different characteristics, and use the classification strategy to assign labels to each cluster after clustering; The specific steps of S2.3 also include: S2.33: Use deep reinforcement learning algorithm to learn the optimal behavior strategy based on the characteristics of the truck in different states, where the cluster label is used as the state space; S2.34: Based on the actual application results, collect feedback information, iteratively optimize the clustering algorithm, classification strategy, and deep reinforcement learning model, and adjust the parameters and structure of the deep reinforcement learning model.

2. The commercial truck image construction and label assignment method under multi-source heterogeneous data fusion according to claim 1 is characterized in that: The tag calculation method in S2.2 specifically includes: S2.21: Determine the label calculation method and value; If it is a fact label, it is assigned a value based on the feature value in the data set and the value range is set , where a and b represent the minimum and maximum values ​​of the fact label value range respectively; If it is an advanced label, it is calculated based on multiple fact labels through weighted summation and the value range is set. , where c and d represent the minimum and maximum values ​​of the advanced tag value range, respectively; If it is a model label, the label value is calculated using regression statistics, and the level is set to high, medium, or low according to the label value; S2.22: Use the defined label calculation method and X to calculate the label value of each feature object, and assign the calculated label value to the corresponding feature object to build a label system.

3. The commercial truck image construction and label assignment method under multi-source heterogeneous data fusion according to claim 2 is characterized in that: The specific formula of the deep reinforcement learning algorithm in S2.33 is: ； in, Representation parameters The gradient of represents the gradient function, Indicated in strategy The expectations below, represents the action at time t, Represents the policy function, which is in the state Next select action The probability of represents the state at time t, represents the advantage function, which is the cumulative reward from time t to the terminal state, represents the discount factor, Indicates status The value function under Indicates status The value function under represents the baseline function, and T represents the time length.

4. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the commercial truck portrait construction and label assignment method under multi-source heterogeneous data fusion described in any one of claims 1 to 3 are implemented.

5. A computer-readable storage medium, characterized in that: Computer instructions are stored thereon, and when the computer instructions are executed, the steps of the commercial truck portrait construction and label assignment method under multi-source heterogeneous data fusion described in any one of claims 1-3 are executed.

Images