Vehicle delivery fault prediction method

Abstract

The invention discloses a vehicle factory failure prediction method, which comprises the following steps of: collecting factory vehicle state parameter data, storing parameter data according to a preset format, extracting the parameter data for preprocessing, converting the parameter data into a standard data format, performing dimension reduction on the preprocessed data to extract related characteristics, and performing model training by adopting a random forest algorithm. The method has the advantages that potential quality problems can be recognized in advance, early warning can be performed in advance, detection efficiency and accuracy are improved, and the like.

Description

Technical Field

[0001] The invention relates to the technical field of special vehicles, in particular to a method for predicting vehicle factory failures. Background Art

[0002] In the modern aviation industry, airport special vehicles such as shuttle buses, baggage trucks, and tractors play a vital role in ensuring efficient and safe airport operations. Traditional vehicle factory inspections mainly rely on manual inspections and regular inspections, which are usually inefficient, difficult to cover all potential failure points, and susceptible to human errors; at the same time, although a large amount of data is generated during the inspection process, this data is often unable to be effectively dealt with, resulting in inefficient data mining and analysis.

[0003] In addition, existing detection methods mainly focus on static detection and lack a predictive maintenance mechanism, which means that potential faults are often not discovered until after the vehicle is put into use, affecting the stability and safety of operations.

[0004] Therefore, there is an urgent need for a technical solution that can intelligently process and analyze detection data, realize factory fault prediction, and improve the accuracy and reliability of detection. Summary of the invention

[0005] The purpose of the present invention is to solve the deficiencies of the above-mentioned prior art and to provide a vehicle factory fault prediction method that can identify potential quality problems in advance, provide early warning, and improve detection efficiency and accuracy.

[0006] The technical solution adopted by the present invention to solve its technical problem is:

[0007] A vehicle factory fault prediction method, characterized by:

[0008] Step S1: Collecting factory vehicle status parameter data and storing the parameter data in a predetermined format;

[0009] Step S2: extract parameter data for preprocessing and convert it into a standard data format;

[0010] Step S3: Perform dimensionality reduction on the preprocessed data to extract relevant features;

[0011] Step S4: using random forest algorithm to train the model;

[0012] Step S5: Use the trained model to predict the faults of the factory vehicles and determine whether the vehicles have potential faults;

[0013] It can predict vehicle factory failures, identify potential quality problems, improve detection efficiency and accuracy, provide early warning of possible failures, optimize vehicle quality management, and ensure the safety and reliability of vehicle operations.

[0014] The present invention also includes step S6, continuously monitoring the predictive performance of the model, determining whether the model has performance degradation by collecting new data and performing performance evaluation, and retraining and optimizing the model when the model shows performance degradation, so as to ensure long-term stability and predictive accuracy and adapt to the changing needs of vehicle factory inspection.

[0015] After the model is trained in step S4, the present invention determines whether the model reaches the expected accuracy. If yes, jump to step S5. If not, adjust the hyperparameters of the model and then re-determine whether the model reaches the expected accuracy.

[0016] The specific method in step S1 of the present invention is:

[0017] Step S1.1: The factory vehicle transmits the acquired state parameter data to the Internet of Vehicles platform, defines the parameter of the i-th vehicle as Xi, defines the fault of the i-th vehicle as Yi, and (Xi, Yi) is a standard sample;

[0018] Step S1.2: Extract the original format status parameter data from the Internet of Vehicles platform for processing, and export the data into a standard csv file. The exported csv file can include the parameter information of all vehicles and the fault set of each vehicle in one table.

[0019] The specific method in step S2 of the present invention is:

[0020] Step S2.1: Use median interpolation method to clean the data. The formula is:

[0021]

[0022] Among them, x i is the original data value, x i ' is the processed data value, Median(X) is the median of feature X;

[0023] Step S2.2: Tag value encoding

[0024] Step S2.2.1: Count the total number of fault types N that appear in the database, and arrange the fault types in a certain order and number them from 1 to N, with N serial numbers corresponding to N fault types one by one, and store them in a dictionary for future query;

[0025] Step S2.2.2: define an N-dimensional row vector yi, where i is the i-th vehicle;

[0026] Step S2.2.3: Query the fault condition of each vehicle. If a vehicle has k faults among N fault types, query the k serial numbers corresponding to these k faults in step S2.2.1, match the numbers corresponding to these k serial numbers in yi with the corresponding dimensions in yi, so that the components of the corresponding dimensions of yi are 1, and the rest are 0;

[0027] Preprocessing the extracted raw data can ensure data quality and model effectiveness. Data cleaning can ensure data integrity and consistency by removing or repairing missing data and processing outliers, eliminating noise and incomplete information that may affect model performance. The median interpolation method can ensure that the data mean remains unchanged while greatly reducing the impact of outliers on the overall data pattern. Label value encoding structures the original multidimensional data into a standardized format to meet the input requirements of the machine learning model. The encoding method is simple and easy to use. Finally, the preprocessed data is organized into a form that can be directly read and analyzed by the algorithm, ensuring that the subsequent algorithm can accurately and efficiently extract useful information from it and make predictions.

[0028] The specific method of using PCA technology to reduce the dimension of data and extract features in step S3 of the present invention is:

[0029] Step S3.1: define the characteristic parameters of special vehicles as a matrix X of dimension n×p, that is, it is assumed that there are n vehicles in the database, and each vehicle has p characteristic parameters;

[0030] Step S3.2: normalize the matrix X to obtain a normalized feature matrix Z;

[0031]

[0032] Among them, μ is the mean vector of each feature, σ is the standard deviation vector of each feature;

[0033] Step S3.3: Calculate the covariance matrix C of the standardized data;

[0034]

[0035] Step S3.4: Calculate the eigenvalues ​​and eigenvectors of C, and select the eigenvectors corresponding to the first k eigenvalues ​​according to the eigenvalue order to form a new k-order eigenmatrix V;

[0036] Step S3.5: the reduced-dimensional data matrix Y=ZV;

[0037] Among them, Y is the principal component matrix after dimensionality reduction, and its dimension is reduced from n×p to n×k, that is, the number of vehicles remains unchanged but only the first k most important features are retained, and pk features with low correlation are eliminated;

[0038] Step S3.6: Rank the variables by importance. The importance of variables in random forest is measured by the reduction of Gini index.

[0039]

[0040] Among them, X j is a certain feature, T is the number of all numbers Gini before and Gini after The Gini index before and after the split is

[0041] PCA technology is used to reduce the dimensionality of high-dimensional data and extract the most representative features while retaining the main information of the data. When the data is very high-dimensional, PCA dimensionality reduction can significantly reduce the computational cost. Random forest is then used for importance analysis to quantify the importance of the features in the vehicle's principal components to the prediction results. If computing resources are tight, this method can be used to further eliminate parameters with lower correlation.

[0042] In step S3 of the present invention, before using the PCA technology to reduce the dimension of the data and extract features, low-correlation parameters are first manually removed from the preprocessed data; due to the excessive number of data features, there are low-correlation parameters that can be directly judged through experience. First, based on experience, some obvious low-correlation parameters are manually removed to ensure that the data set covers parameters that have an important impact on vehicle fault prediction. The remaining features are then processed by PCA technology, which can improve the accuracy and efficiency of the algorithm and prevent low-correlation parameters from interfering with PCA.

[0043] The specific method of model training in step S4 of the present invention is:

[0044] Step S4.1: Divide the data set into a training set and a validation set, and use the training set to build a decision tree model;

[0045] Step S4.2: Determine the final classification result by letting the prediction results of all decision trees take majority vote;

[0046] Step S4.3: Hyperparameter tuning: Use grid search and cross-validation methods to adjust the model's hyperparameters to optimize model performance. Grid search traverses parameter combinations and selects the parameters that perform best on the validation set. The formula for hyperparameter tuning is:

[0047]

[0048] Among them, Θ represents the set of all parameter combinations, and θ is one of the parameter combinations;

[0049] Step S4.4: Save the parameter combination that performs best in the validation set, and the model training is completed;

[0050] Step S4.5: Evaluate the trained model. Common evaluation criteria include accuracy, precision, recall, and F1 value. These evaluation criteria should be combined with k-fold cross validation. The formula for k-fold cross validation is as follows:

[0051]

[0052] Where Accuracyi is the accuracy of the i-th fold;

[0053] The calculation formulas for each indicator are as follows:

[0054]

[0055] Where TP is a true positive example, TN is a true negative example, FP is a false positive example, and FN is a false negative example;

[0056] This algorithm can improve the robustness and prediction accuracy of the model by constructing multiple decision trees and combining their prediction results, thereby enhancing the ability to accurately predict the vehicle's factory status.

[0057] The method for constructing the decision tree in step S4.1 of the present invention is:

[0058] Step S4.1.1: Calculate the Gini index of all vehicle corresponding samples. Assuming there are n samples, calculate the Gini index G(X1), G(X2), ... G(Xn) of these n samples respectively. The definition of Gini index is as follows:

[0059]

[0060] Among them, p k is the proportion of samples belonging to the kth category, K is the total number of categories;

[0061] Step S4.1.2: For each feature, calculate its weighted Gini index when splitting the data set. The calculation of the weighted Gini index considers each value of the feature and its corresponding subset:

[0062]

[0063] Where T is the data set of the current node, X is the feature, Ti is the data subset when feature X takes the i-th value, |Ti| is the number of samples in subset Ti, |T| is the number of samples in the data set T of the current node, and G(Ti) is the Gini index of subset Ti;

[0064] Step S4.1.3: Select the feature with the smallest weighted Gini index as the first splitting feature of the root node. After the data is split, it is divided into different subsets. Recursively apply the same process to each subset, continue to select the feature with the smallest Gini index for further splitting until the stopping condition is met;

[0065] By calculating the weighted Gini index of all features and selecting the feature with the smallest Gini index for data splitting, the feature with the smallest Gini index can minimize the uncertainty of the data set to the greatest extent. Therefore, it is considered the best splitting feature.

[0066] In step S4 of the present invention, the training data of each decision tree is constructed by Bootstrap sampling, where a part of the samples are selected each time with replacement. To predict whether a certain vehicle fault yi will occur, first draw m samples from n samples with replacement, where m < n, to obtain a subset of all samples. Repeat this process to obtain p subsets. p different decision trees can be constructed using the p subsets, and they all give the predicted value of a certain fault yi;

[0067] Suppose more than half of the p decision trees think the value of yi is 0, and the remaining ones think the value of yi is 1. Then, according to the voting result, the random forest gives the final predicted value of yi as 0, that is, it is considered that the fault will not occur. The output result of the random forest is as shown in the formula:

[0068]

[0069] In the formula, B represents the number of decision trees, and each decision tree h b (x) gives a classification prediction result, and ck represents a certain category.

[0070] The beneficial effects of the present invention are as follows: It can realize the prediction of vehicle factory faults, identify potential quality problems, improve the detection efficiency and accuracy, give early warnings of possible faults, optimize the quality management of vehicles, and ensure the safety and reliability of vehicle operation. Description of the Drawings

[0071] Figure 1 is the flowchart of the prediction of factory faults of airport special vehicles in the present invention.

[0072] Figure 2 are some parameters of the vehicle collected in the present invention.

[0073] Figure 3 is the original data format exported from the vehicle networking platform in the present invention.

[0074] Figure 4 is the style of the csv file exported in the present invention.

[0075] Figure 5It is a schematic diagram of constructing a decision tree of the present invention. DETAILED DESCRIPTION

[0076] The present invention is described below in conjunction with the accompanying drawings and embodiments.

[0077] As attached Figure 1 As shown, a vehicle factory fault prediction method,

[0078] Step S1: Collecting factory vehicle status parameter data and storing the parameter data in a predetermined format;

[0079] Step S1.1: The factory vehicle transmits the acquired status parameter data to the Internet of Vehicles platform for real-time monitoring and subsequent data analysis. The collected parameter data include the detection data obtained from various sensors of airport special vehicles (the vehicle is equipped with multiple sensors, such as speed sensor, throttle sensor, brake sensor, temperature sensor, etc.), performance test equipment and maintenance record system. These data cover various performance indicators and historical operating status of the vehicle, including but not limited to sensor measurement data, component function test results and maintenance history records. Figure 2 Some relevant parameters collected are shown;

[0080] like Figure 2 As shown in the figure, the parameters of the i-th vehicle in the Internet of Vehicles platform are defined as Xi, Xi is a row vector, and contains all the relevant parameters of the i-th vehicle collected. The fault of the i-th vehicle is defined as Yi, Yi is also a row vector, where the value range of i changes dynamically with the increase or decrease of the number of vehicles in the Internet of Vehicles platform. (Xi, Yi) is a standard sample.

[0081] Step S1.2: extract the original format state parameter data from the Internet of Vehicles platform for processing. Figure 3 The original data format is shown. The parameter name and the corresponding specific value are usually not in the same table. When querying, you need to search in another database by id. Export the data to a standard csv file to facilitate the storage of structured data. Each row is a sample, and the columns are the parameter data Xi of each state. The last column is the label value Yi. The exported csv file can include the parameter information of all vehicles and the fault set of each vehicle in one table. The set (Xi, Yi) constitutes a sample. Figure 4 Shows the export standard csv format.

[0082] Step S2: extract parameter data for preprocessing and convert it into a standard data format that can be read by the algorithm;

[0083] Step S2.1: Use median interpolation method to clean the data. The formula is:

[0084]

[0085] Among them, x i is the original data value, x i ' is the processed data value, Median(X) is the median of feature X;

[0086] by Figure 4 Take the data in as an example, there are two null values. If you read the data directly at this time, the program will report an error. Because the data is abnormal, it is necessary to clean the data. Using median interpolation, you can use Excel to calculate the median of the normal values ​​in the "motor speed" and "S2 pressure" columns respectively, set them as m and n, and replace the two null values ​​with m and n respectively, which can eliminate the abnormal values ​​in the data, ensure the integrity and coherence of the data, and thus not affect the normal operation of the algorithm program;

[0087] Step S2.2: Tag value encoding

[0088] Step S2.2.1: Count the total number of fault types N that appear in the database, and arrange the fault types in a certain order and number them from 1 to N, with N serial numbers corresponding to N fault types one by one, and store them in a dictionary for future query;

[0089] Step S2.2.2: define an N-dimensional row vector yi (i is the i-th vehicle);

[0090] Step S2.2.3: Query the fault condition of each vehicle. If a vehicle has k faults among N fault types, query the k serial numbers corresponding to these k faults in step S2.2.1, match the numbers corresponding to these k serial numbers in yi with the corresponding dimensions in yi, so that the components of the corresponding dimensions of yi are 1, and the rest are 0;

[0091] For example, there are 10 types of vehicle faults in the database, namely (1-accelerator pedal disconnection, 2-accumulator pressure sensor short circuit, 3-hydraulic bypass valve output fault, 4-low-voltage battery fault, 5-left front encoder fault, 6-instrument fault, 7-power battery fault, 8-air conditioning disconnection, 9-high-voltage distribution system fault, 10-low coolant level). After query, the j-th vehicle in the database has faults corresponding to serial numbers 1, 3, 5, 7, and 9. Define a 10-dimensional vector yj, assign 1 to the 1st, 3rd, 5th, 7th, and 9th dimensions in yj, and the rest are 0. Then the fault information of the j-th special vehicle is converted from (accelerator pedal disconnection, hydraulic bypass valve output fault, left front encoder fault, power battery fault, high-voltage distribution system fault) to yj: (1, 0, 1, 0, 1, 0, 1, 0), and the encoding method is simple and easy to use, and can be recognized by the random forest algorithm;

[0092] Preprocessing the extracted raw data can ensure data quality and model effectiveness. Data cleaning can ensure data integrity and consistency by removing or repairing missing data and processing outliers, eliminating noise and incomplete information that may affect model performance. The median interpolation method can ensure that the data mean remains unchanged while greatly reducing the impact of outliers on the overall data pattern. Label value encoding structures the original multidimensional data into a standardized format to meet the input requirements of the machine learning model. The encoding method is simple and easy to use. Finally, the preprocessed data is organized into a form that can be directly read and analyzed by the algorithm, ensuring that the subsequent algorithm can accurately and efficiently extract useful information from it and make predictions.

[0093] Step S3: Perform dimensionality reduction on the preprocessed data to extract relevant features;

[0094] Step S3.1: First, manually remove low-correlation parameters from the preprocessed data. Due to the excessive number of data features, there are low-correlation parameters that can be directly judged by experience. First, based on experience, some obvious low-correlation parameters are manually removed to ensure that the data set covers parameters that have an important impact on vehicle fault prediction, and extract features related to vehicle performance and operating status, such as engine temperature, oil pressure, speed sensor, battery voltage, etc. The extracted features are then processed by PCA technology to improve the accuracy and efficiency of the algorithm and prevent low-correlation parameters from interfering with PCA.

[0095] Step S3.2: PCA technology is used to reduce the dimension of the data and extract features. The characteristic parameters of special vehicles are defined as a matrix X with dimension n×p, that is, it is assumed that there are n vehicles in the database and each vehicle has p characteristic parameters;

[0096] Step S3.3: Standardize the matrix X to eliminate the influence of different feature dimensions and obtain a standardized feature matrix Z;

[0097]

[0098] Among them, μ is the mean vector of each feature, σ is the standard deviation vector of each feature;

[0099] Step S3.4: Calculate the covariance matrix C of the standardized data;

[0100]

[0101] Step S3.5: Calculate the eigenvalues ​​and eigenvectors of C, and select the eigenvectors corresponding to the first k eigenvalues ​​according to the eigenvalue order to form a new k-order eigenmatrix V;

[0102] Step S3.6: The reduced-dimensional data matrix Y=ZV;

[0103] Among them, Y is the principal component matrix after dimensionality reduction, and its dimension is reduced from n×p to n×k, that is, the number of vehicles remains unchanged but only the first k most important features are retained, and pk features with low correlation are eliminated;

[0104] Step S3.7: Rank the variables by importance. The importance of variables in random forest is measured by the reduction of Gini index.

[0105]

[0106] Among them, X j is a certain feature, T is the number of all numbers Gini before and Gini after The Gini index before and after the split is

[0107] PCA technology is used to reduce the dimensionality of high-dimensional data and extract the most representative features while retaining the main information of the data. When the data is very high-dimensional, PCA dimensionality reduction can significantly reduce the computational cost. Random forest is then used for importance analysis to quantify the importance of the features in the vehicle's principal components to the prediction results. If computing resources are tight, this method can be used to further eliminate parameters with lower correlation.

[0108] Step S4: Use the improved random forest algorithm to train the model on the vehicle's factory inspection data. The specific method of model training is as follows:

[0109] Step S4.1: Divide the data set into a training set and a validation set, and use the training set to build a decision tree model;

[0110] Step S4.1.1: Use the Gini index to construct split features and calculate the Gini index of all vehicle corresponding samples. Assuming there are n samples, calculate the Gini index G(X1), G(X2),...G(Xn) of these n samples respectively. The definition of the Gini index is as follows:

[0111]

[0112] Among them, p k is the proportion of samples belonging to the kth category, K is the total number of categories;

[0113] Step S4.1.2: For each feature, calculate its weighted Gini index when splitting the data set. The calculation of the weighted Gini index considers each value of the feature and its corresponding subset:

[0114]

[0115] Among them, T is the data set of the current node, X is the feature, Ti is the data subset when the feature X takes the i-th value, |Ti| is the number of samples in the subset Ti, |T| is the number of samples in the current node data set T, and G(Ti) is the Gini index of the subset Ti;

[0116] Step S4.1.3: Select the feature with the smallest weighted Gini index as the first splitting feature of the root node. After the data is split, it is divided into different subsets. Apply the same process recursively on each subset, and continue to select the feature with the smallest Gini index for further splitting until the stopping condition is met, such as the maximum depth of the tree, the number of samples in the node is lower than a certain threshold, or the Gini index of the node is lower than a certain threshold. Thus, the construction of a decision tree is completed;

[0117] By calculating the weighted Gini index of all features and selecting the feature with the smallest Gini index for data splitting, the feature with the smallest Gini index can minimize the uncertainty of the data set to the greatest extent. Therefore, it is considered the best splitting feature;

[0118] Step S4.2: Determine the final classification result by having the prediction results of all decision trees vote majority;

[0119] For example, to predict whether a certain vehicle fault yi will occur, first use Bootstrap sampling to draw m samples (m < n) with replacement from n samples to obtain a subset of the entire sample. Repeat this process to obtain p subsets. p different decision trees can be constructed using these p subsets, and they all give the predicted value of a certain fault yi. Suppose more than half of the p decision trees think the value of yi is 0, and the remaining ones think the value of yi is 1. Then, according to the voting result, the random forest gives the final predicted value of yi as 0, that is, it is considered that the fault will not occur. The output result of the random forest is as shown in the formula:

[0120]

[0121] In the formula, B represents the number of decision trees, and each decision tree h b (x) will give a classification prediction result, and ck represents a certain category;

[0122] Step S4.3: Hyperparameter tuning. Use the grid search and cross-validation methods to adjust the hyperparameters of the model to optimize the model performance. The grid search traverses the parameter combinations and selects the parameters that perform best on the validation set. The formula for hyperparameter tuning is:

[0123]

[0124] Among them, Θ represents the set of all parameter combinations, and θ is one of the parameter combinations;

[0125] Step S4.4: Save the parameter combination that performs best in the validation set, and the model training is completed;

[0126] Step S4.5: Evaluate the trained model. Common evaluation criteria include accuracy, precision, recall, and F1 value. These evaluation criteria should be combined with k-fold cross validation. The formula for k-fold cross validation is as follows:

[0127]

[0128] Where Accuracyi is the accuracy of the i-th fold;

[0129] The calculation formulas for each indicator are as follows:

[0130]

[0131] Where TP is a true positive example, TN is a true negative example, FP is a false positive example, and FN is a false negative example;

[0132] The k-fold cross-validation method is that the device has N samples, the N samples are averaged and randomly divided into k equal parts, the first sample is used as the validation set, and the remaining k-1 samples are used as training sets to calculate the accuracy once, then the second sample is used as the validation set, and the remaining k-1 samples are used as training sets to calculate the accuracy again, until every sample is used as a validation set to participate in the calculation, and the k-fold accuracy obtained is averaged as the cross-validation accuracy;

[0133] Taking 5-fold cross validation as an example, there are 100 samples of special vehicles. The 100 samples are averaged and randomly divided into five equal parts. The first sample is used as a validation set, and the remaining four samples are used as training sets to calculate the accuracy once. Next, the second sample is used as a validation set, and the remaining four samples are used as training sets to calculate the accuracy again. Until each sample has participated in the calculation as a validation set, the average of the five accuracy rates is taken as the cross-validation accuracy.

[0134] This algorithm can improve the robustness and prediction accuracy of the model by constructing multiple decision trees and combining their prediction results, thereby improving the ability to accurately predict the vehicle's factory status;

[0135] After the model is trained, determine whether the model has achieved the expected accuracy. If so, jump to step S5. If not, adjust the hyperparameters of the model and then re-determine whether the model has achieved the expected accuracy.

[0136] Step S5: Use the trained improved random forest model to predict the faults of the factory vehicles and determine whether the vehicles have potential faults. This can predict the factory faults of the vehicles. The prediction results are sent to the client through the Internet of Vehicles platform for engineers to view in real time, identify potential quality problems, improve detection efficiency and accuracy, and warn of possible faults in advance. Take necessary measures to reduce the risks of vehicles in actual operation, optimize vehicle quality management, and ensure the safety and reliability of vehicle operations.

[0137] Step S6, continuously monitor and improve the prediction performance of the random forest model, and determine whether the model has performance degradation by collecting new data and performing performance evaluation. When the model shows a decrease in accuracy or insufficient adaptability, retrain and optimize the model to ensure long-term stability and prediction accuracy, and adapt to the changing needs of vehicle factory inspection.

[0138] The above steps can predict vehicle factory failures, identify potential quality problems, improve detection efficiency and accuracy, provide early warning of possible failures, optimize vehicle quality management, and ensure the safety and reliability of vehicle operations.

[0139] This application is not only applicable to the factory fault prediction of special vehicles at airports, but also to vehicles used in other scenarios.

Claims

1. A vehicle factory fault prediction method, characterized by: Step S1: Collecting factory vehicle status parameter data and storing the parameter data in a predetermined format; Step S2: extract parameter data for preprocessing and convert it into a standard data format; Step S3: Perform dimensionality reduction on the preprocessed data to extract relevant features; Step S4: using random forest algorithm to train the model; Step S5: Use the trained model to predict faults of the factory vehicles and determine whether the vehicles have potential faults.

2. A vehicle factory failure prediction method according to claim 1, characterized in that: The method also includes step S6, which continuously monitors the prediction performance of the model, determines whether the model has performance degradation by collecting new data and performing performance evaluation, and retrains and optimizes the model when the model shows performance degradation.

3. A vehicle factory failure prediction method according to claim 1 or 2, characterized in that: After the model is trained in step S4, it is determined whether the model has reached the expected accuracy. If so, the process jumps to step S5. If not, the hyperparameters of the model are adjusted and then it is determined again whether the model has reached the expected accuracy.

4. A vehicle factory failure prediction method according to claim 1 or 2, characterized in that: The specific method in step S1 is: Step S1.1: The factory vehicle transmits the acquired state parameter data to the Internet of Vehicles platform, defines the parameter of the i-th vehicle as Xi, defines the fault of the i-th vehicle as Yi, and (Xi, Yi) is a standard sample; Step S1.2: Extract the original format status parameter data from the Internet of Vehicles platform for processing, and export the data into a standard csv file. The exported csv file can include the parameter information of all vehicles and the fault set of each vehicle in one table.

5. A vehicle factory failure prediction method according to claim 1 or 2, characterized in that: The specific method in step S2 is: Step S2.1: Use median interpolation method to clean the data. The formula is: Among them, x i is the original data value, x i ' is the processed data value, Median(X) is the median of feature X; Step S2.2: Tag value encoding Step S2.2.1: Count the total number of fault types N that appear in the database, and arrange the fault types in a certain order and number them from 1 to N, with N serial numbers corresponding to N fault types one by one, and store them in a dictionary for future query; Step S2.2.2: define an N-dimensional row vector yi, where i is the i-th vehicle; Step S2.2.3: Query the fault condition of each vehicle. If a vehicle has k types of faults out of N fault types, query the k serial numbers corresponding to these k types of faults in step S2.2.1, and match the numbers corresponding to these k serial numbers in yi with the corresponding dimensions in yi, so that the components of the corresponding dimensions of yi are 1, and the rest are 0.

6. A vehicle factory failure prediction method according to claim 1 or 2, characterized in that: In step S3, the PCA technology is used to reduce the dimension of the data and extract features. The specific method is: Step S3.1: define the characteristic parameters of special vehicles as a matrix X of dimension n×p, that is, it is assumed that there are n vehicles in the database, and each vehicle has p characteristic parameters; Step S3.2: normalize the matrix X to obtain a normalized feature matrix Z; Among them, μ is the mean vector of each feature, σ is the standard deviation vector of each feature; Step S3.3: Calculate the covariance matrix C of the standardized data; Step S3.4: Calculate the eigenvalues ​​and eigenvectors of C, and select the eigenvectors corresponding to the first k eigenvalues ​​according to the eigenvalue order to form a new k-order eigenmatrix V; Step S3.5: the reduced-dimensional data matrix Y=ZV; Among them, Y is the principal component matrix after dimensionality reduction, and its dimension is reduced from n×p to n×k, that is, the number of vehicles remains unchanged but only the first k most important features are retained, and pk features with low correlation are eliminated; Step S3.6: Rank the variables by importance. The importance of variables in random forest is measured by the reduction of Gini index.

7. A vehicle factory failure prediction method according to claim 6, characterized in that: In step S3, before using the PCA technology to reduce the dimension of the data and extract features, low-correlation parameters are manually removed from the pre-processed data.

8. A vehicle factory failure prediction method according to claim 1, 2 or 7, characterized in that: The specific method of model training in step S4 is: Step S4.1: Divide the data set into a training set and a validation set, and use the training set to build a decision tree model; Step S4.2: Determine the final classification result by letting the prediction results of all decision trees take majority vote; Step S4.3: Hyperparameter tuning: Use grid search and cross-validation methods to adjust the model's hyperparameters to optimize model performance. Grid search traverses parameter combinations and selects the parameters that perform best on the validation set. The formula for hyperparameter tuning is: Among them, Θ represents the set of all parameter combinations, and θ is one of the parameter combinations; Step S4.4: Save the parameter combination that performs best in the validation set, and the model training is completed; Step S4.5: Evaluate the trained model. Common evaluation criteria include accuracy, precision, recall, and F1 value. These evaluation criteria should be combined with k-fold cross validation. The formula for k-fold cross validation is as follows: Where Accuracyi is the accuracy of the i-th fold; The calculation formulas for each indicator are as follows: Where TP is a true positive example, TN is a true negative example, FP is a false positive example, and FN is a false negative example.

9. A vehicle factory failure prediction method according to claim 8, characterized in that: The decision tree construction method in step S4.1 is: Step S4.1.1: Calculate the Gini index of all vehicle corresponding samples. Assuming there are n samples, calculate the Gini index G(X1), G(X2), ... G(Xn) of these n samples respectively. The definition of Gini index is as follows: Among them, p k is the proportion of samples belonging to the kth category, K is the total number of categories; Step S4.1.2: For each feature, calculate its weighted Gini index when splitting the data set. The calculation of the weighted Gini index considers each value of the feature and its corresponding subset: Where T is the data set of the current node, X is the feature, Ti is the data subset when feature X takes the i-th value, |Ti| is the number of samples in subset Ti, |T| is the number of samples in the data set T of the current node, and G(Ti) is the Gini index of subset Ti; Step S4.1.3: Select the feature with the smallest weighted Gini index as the first split feature of the root node. After the data is split, it is divided into different subsets. The same process is recursively applied to each subset, and the feature with the smallest Gini index is continued to be selected for further splitting until the stopping condition is met.

10. A vehicle factory failure prediction method according to claim 8, characterized in that: In step S4, the training data of each decision tree is constructed by Bootstrap sampling, where a part of the samples are selected each time with replacement. To predict whether a certain vehicle fault yi will occur, first, m samples are drawn from n samples with replacement (m < n) to obtain a subset of all samples. This process is repeated p times to obtain p subsets. Using these p subsets, p different decision trees can be constructed, and they all give the predicted value of a certain fault yi; If more than half of the p decision trees think the value of yi is 0 and the remaining ones think the value of yi is 1, then according to the voting result, the random forest gives the final predicted value of yi as 0, that is, it is considered that the fault will not occur. The output result of the random forest is as shown in the formula: In the formula, B represents the number of decision trees, and each decision tree h b (x) will give a classification prediction result, ck represents a certain category.

Images